Web par General Agents: Aik Crash Course
12 Concepts · Aap ka pehla general agent, usi browser tab mein jahan aap ne Foundations ki thi
Chhe courses tak aap chat tab mein kaam karte rahe. Aap ne poocha, us ne jawab diya. Aap ne phir poocha. Har turn aap ne shuru kiya. Seekhne ka yeh acha tareeqa tha. Lekin plain chat ki aik limit hai: aap type karna band karte hain to conversation bhi ruk jati hai.
Apne aap kaam jari rakhne wale agents web par naye nahin. ChatGPT 2025 ke shuru se scheduled tasks chala sakta tha. Aap ke door jane par woh apne aap deep research kar sakta tha. Us ke paas apne computer wala agent mode bhi tha. Yeh capabilities asli aur powerful thin, lekin bikhri hui: aik task yahan, research tool wahan, agent mode kisi menu mein. Rozmarra kaam ke liye koi aik saaf jagah nahin thi jahan aap ki files, sessions aur services aik flow mein milen.
July 2026 ne yeh badal diya. Do bari companies ne in pieces ko chat box ke paas, usi web address par, aik saaf work surface mein jor diya. claude.ai par Claude Cowork Anthropic ke servers par remote session ki tarah chalta hai. chatgpt.com par ChatGPT Work, ChatGPT ke andar aik naya mode hai, usi Chat mode ke paas jisay aap pehle se jante hain. Yeh OpenAI ke GPT-5.6 model par chalta hai. (ChatGPT mein coders ke liye Codex mode bhi hai, lekin web aur phone par zyada tar Chat aur Work dikhte hain.)
Is liye July 2026 ne browser agents invent nahin kiye. Us ne unhein kaam ki first-class jagah di. Chat box nahin badla. Woh ab bhi aap ka intezar karta hai aur hamesha karega. Us ke paas wala surface badla hai. Wahan session shuru karein to woh company ke servers par chalta hai. Tab band kar sakte hain. Plane mein baith sakte hain. Kaam jari rehta hai. Jab aisi choice aati hai jo sirf aap kar sakte hain, sawaal aap ke phone par aata hai.
Yeh course us surface ka tour hai. Saath hi, yeh us skill ka aap ka pehla lesson hai jis ki har baad wale course ko zaroorat hai: general agent ko direct karna. Aap seekhenge remote session kya hai, files asal mein kahan jati hain, approval phone tak kaise pohanchti hai aur us kaam ko hand off kaise karna hai jisay aap dekhte nahin rahenge. Kisi naye agent product ko dekh kar us mein wohi parts pehchanna bhi seekhenge.
Do companies ne taqreeban aik hi waqt par isay banaya aur dono usi set of parts tak pohanchin. Yeh is kitab ke bunyadi idea ka mazboot saboot hai: skill tool mein nahin, shape mein hai.
Yahan aik bari kahani bhi hai jisay shuru mein naam dena zaroori hai. Dono companies ab sirf model par race nahin kar rahin. Woh model ke gird har cheez par race kar rahi hain: schedule, connectors, memory aur gate. Yeh kitab us wrapper ko harness aur us race ko harness war kehti hai. Pehle engineers code mein yeh race larte the. Ab yeh web tak aa gayi hai. Is mein aap ki seat seedhi hai: yahan harness nahin banate. Jo fit ho usay chunte aur achi tarah direct karte hain. Concept 3 dono harnesses ko side by side dikhata hai.
Aap kuch install nahin karenge. Foundations browser tab mein chali thi. Yeh bhi wahin chalta hai.
Yeh course kis ke liye hai. Foundations se aane wale har us shakhs ke liye jo kuch install kiye baghair real agent direct karna chahta hai. Sab se pehle knowledge workers: jin ka din documents, spreadsheets, email aur chat mein guzarta hai. Lawyers, accountants, marketers, analysts, founders aur students. Un engineers ke liye bhi jo terminal se pehle sab se narm pehla qadam chahte hain. Code aata hai to yahan se shuru karne mein kuch nahin khoyenge. Code nahin aata to yehi sahi jagah hai. Isi liye yeh General Agents section ka pehla course hai. Har shakhs pehle din browser mein isay kar sakta hai.
Pehle yeh karein: Foundations, khas taur par 2026 mein AI Prompting aur Skills & Connectors. Yeh course aap se teen cheezein expect karta hai: model ko saaf brief karna, context ka matlab janna, aur Skill aur Connector samajhna. In mein se kuch naya hai to pehle woh do courses karein. Dono mil kar taqreeban do ghante lete hain.
Kitna waqt lagega. Core route (Parts 1-3) parhne mein taqreeban 90 minutes. Part 4 aur appendices ke saath poora course qareeban do ghante. Projects ka hands-on waqt alag hai, jis mein scheduling exercise ka aik hafta shamil hai, aur learning wahin mazboot hoti hai. Aage reference ke liye saat appendices hain: chhe is course ko aage ke courses se map karte hain (jab wahan pohanchein tab parhein), aur aakhri poori industry ko map karta hai. Inhein reference ki tarah parhein, aik sitting mein nahin.
Naye hain? Jo pehle se aana chahiye us ka 2-minute recap
- Briefing ka matlab task ko mukammal describe karna hai: outcome, limits, reader aur reason. Vague prompt se vague result milta hai. (Foundations Course 2.)
- Context woh sab hai jo model aik turn mein dekhta hai. Zyada hamesha behtar nahin hota. Sahi context ko maqsad ke saath chunna behtar hai. (Foundations Course 2.)
- Skill woh task hai jo aap aik baar sikhate hain. Yeh
SKILL.mdfile mein rehta hai. AI isay tabhi load karta hai jab request match karti hai. (Foundations Course 5.) - Connector MCP standard ke zariye AI ko Drive, Gmail ya Slack jaisi real service tak mehfooz, mahdood access deta hai. Permissions aap dete hain. AI unhi ke andar kaam karta hai. (Foundations Course 5.)
- Chatbot aur agent. Chatbot jawab deta hai. Agent plan karta hai, tools use karta aur qadam mukammal karta hai. Is course mein aap pehli baar aik agent direct karenge.
In mein se kuch naya hai to pehle woh do Foundations courses parhein. Yeh course in sab ko real work mein use karta hai.
📚 Teaching Aid
Poori Presentation Dekhein - Web par General Agents
Instructors aur self-learners ke liye. Yeh course aik transferable idea sikhata hai: agent product ke chhe parts hote hain (heartbeat, connectors, run-until-done loop, state spine, human gate, body). Parts samajhne ke baad koi naya product naya subject nahin, sirf chhota look-around ban jata hai. Isay teen passes mein sikhayein. Pehla, shift: chat box intezar karta hai, paas wala agent surface nahin (Concepts 1-3). Doosra, surface: files kahan rehti hain, connectors bahar kaise pohanchte hain, approvals phone tak kaise aati hain (Concepts 4-7). Teesra, unwatched work: delegation loop aur scheduled tasks (Concepts 8-9), phir work kya touch karta hai us ke mutabiq surface chunna (Concepts 10-12). Chhe projects assessment hain. Students ko sab se aham habit de kar bhejein: finished work platform se bahar jata hai.
Seedhi zaban mein aham alfaaz
Yeh alfaaz poore course mein ayenge. List ab aik baar parhein, phir koi term unclear lage to laut aayein.
| Term | Seedha matlab |
|---|---|
| Chat box | Woh conversation jo har turn aap ka intezar karti hai. Apne aap kuch shuru nahin kar sakti |
| Agent surface | Chat box ke paas wala product jo plan karta, action leta aur multi-step work chalata hai |
| Remote session | Vendor ke servers par chalne wala session. Aap ka tab usay dekhne ki window hai, chalane wali machine nahin |
| Local session | Aap ki apni machine par chalne wala session. Machine sleep kare to yeh rukta hai. Agle courses mein milenge |
| Task filesystem | Remote session ke kaam ki temporary jagah. Isay scratch samjhein, permanent nahin: retention aur cleanup rules vendor ke mutabiq badalte hain |
| Platform storage | Vendor platform par aap ke account mein permanently saved files. Task aur tab ke baad bhi baqi rehti hain |
| Exit (Tier 3) | Deliverable ko platform se us jagah le jana jisay aap control karte hain: connector save, download, local write ya repo commit. File aap ke system of record mein aati hai |
| Account spine | Aap ke saved sessions aur files jo devices ke darmiyan sync hote hain. Aisi memory jo aap ko banani nahin pari |
| Human gate | Woh point jahan risky action jari hone se pehle insan ki approval zaroori hai. Is surface par yeh phone tak aati hai |
| Scheduled task | Vendor servers par clock ke mutabiq shuru hone wala kaam, jab aap ka koi device online nahin hota |
| Metered usage | Har plan mein runs ya hours ka budget. Kisi aur ke servers par unattended work ka hamesha aisa budget hota hai |
| Custody | Aap ka data kaun hold karta hai: woh kis ki machines par rehta aur kaun us tak pohanch sakta hai |
Yeh kahan se aaya
Yahan dates aham hain, kyun ke surface chand hafton purana hai lekin is ki roots purani hain.
Browser agents 2026 mein shuru nahin hue. ChatGPT ne January 2025 mein scheduled tasks ship kiye. February 2025 mein deep research ship ki, jo kai minutes apne aap browse karti hai. Mid-2025 mein apne computer wala agent mode ship kiya. Yani is course se bahut pehle web apne aap work chala sakta tha. Bas us work ka aik saaf home nahin tha. Pieces menus aur tools mein phaile the.
Claude apne raste se usi jagah pohancha. Anthropic ne apne servers par chalne wale cloud coding sessions diye jo laptop band hone par bhi chalte the. Mufeed, lekin coders ke liye.
Phir July 2026 mein dono companies ne chand din ke faasle se in pieces ko aik jagah jor diya.
Cowork claude.ai par web mein aaya (beta, pehle Max plan, phir aur plans). Sessions Anthropic servers par chalte hain. Sessions aur files account mein save hote hain. Koi device online na ho tab bhi scheduled tasks fire hote hain.
ChatGPT Work 9 July 2026 ko ship hua. Yeh alag product nahin. ChatGPT ke andar Chat mode ke paas naya agent mode hai aur GPT-5.6 par chalta hai. Yeh finished files banata hai: spreadsheets, slide decks, documents, dashboards aur chhote web apps. Renamed Plugin Directory se aap ki services tak pohanchta hai. Scheduled Tasks aik baar, schedule par, event par ya change monitor karte hue chal sakte hain. Yeh nai price tier nahin. Web aur mobile par eligible paid plans ke saath milta hai, Free ya Go mein nahin, aur har plan ka agent allowance use karta hai. (ChatGPT desktop app mein Chat, Work aur Codex har plan mein hain, Free samet. Codex us desktop app ka alag view hai; web ya mobile par selectable nahin.) Availability badalti rehti hai, is liye live product page par access confirm karein.
Do companies ne chand din ke faasle se chat box ke paas agent rakha. Do products, aik shape. Yeh course shape sikhata hai.
Yeh kyun hua aur ab kya mumkin hai
Surface seekhne se pehle yeh samajhna mufeed hai ke woh maujood kyun hai. Is ke do sides hain: aap ko kya milta hai aur companies ko kya milta hai. Dono parhein. Company kya chahti hai yeh samajh kar blind trust ke bajaye khuli ankhon se surface use karenge. Is se woh rule bhi samajh aata hai jo course aage baar baar dohrayega.
Woh gap jo is ne band kiya. July 2026 se pehle web apne aap work chala sakta tha, lekin pieces bikhre aur uneven the. Scheduled task aik menu mein, research tool doosre mein, agent mode dropdown mein. Har capability akeli powerful thi, lekin saath dhoondna aur combine karna mushkil tha. Aisi koi jagah nahin thi jahan files, sessions, services aur hands-off run milen. Is liye zyada tar logon ne kisi capability ko use hi nahin kiya. In launches ne pieces ko usi address par aik clear surface mein jora jisay log roz kholte hain. Power pehle maujood thi. Ab us tak pohanchna aur combine karna asaan hua.
July 2026 ne aik clear work surface mein chaar capabilities jorin:
- Usi everyday surface mein long-running work. Remote session company ke servers par chalta hai, is liye tab band karne, laptop band karne ya plane mein baithne par bhi task jari rehta hai. Remote agent runs pehle bhi the. Tabdeeli yeh hai ke Work aur Cowork ne unhein ordinary chat ke paas aik general-purpose flow mein rakh diya. (Concept 2.)
- Work surface mein built-in scheduling. Scheduled task clock ya trigger par shuru hota hai, jab aap ka device online nahin. Monday brief aap ke baithne se pehle tayyar hai. Scheduled tasks pehle bhi the. Ab woh alag tool ke bajaye projects, files, connectors aur agent runs ke paas hain. (Concept 9.)
- Sessions aur projects ke gird organized work. Sessions aur files account mein rehte aur devices ke darmiyan aap ke saath chalte hain. Naye surfaces related files, instructions aur continuing work ko aik visible jagah mein rakhte hain, is liye context baar baar paste nahin karna parta. (Concept 4.)
- Files aur connected services mein kai qadam. Agent services se connect hota, plan banata, kai ghanton tak qadam mukammal karta aur chat reply ke bajaye finished files lautata hai. Jin choices ke liye aap chahiye woh phone par aati hain. Naya yeh hai ke sab kuch kai alag product modes ke darmiyan ghoomay baghair hota hai. (Concepts 5-8.)
Ab doosra side dekhein, kyun ke yeh tay karega capabilities ko kaise use karna hai. Zahir kiya goal aap ko zyada kaam mukammal karne ki power dena hai. Business goal reach aur lock-in hai. Poore 2026 mein dono companies is kahani se larti rahin ke woh aise market mein chat product bech rahi hain jahan agents chahiye. Sab se tez fix agent ko wahin rakhna hai jahan karoron log pehle se hain. Koi install nahin, naya app nahin, terminal nahin. Aik move mein chat users agent users ban jate hain. Sessions, files aur connectors ko company account mein save karne ka doosra asar bhi hai: chhorna mehnga ho jata hai. Isay switching cost kehte hain. Kaam ki account memory hi walk away mushkil banati hai.
Yeh shak karne ki wajah nahin, aik saaf habit banane ki wajah hai. Jo feature aap ko sab se zyada madad deta hai (free working memory), wohi company ko sab se zyada madad deta hai (aap ka kaam un ke servers par jama hota hai). Is liye Concept 5 finished work ko platform se bahar apni storage mein bhejna sikhata hai. Convenience rakhein, lock-in se bachein. Company ka goal dikhne ke baad yeh habit paranoid nahin, wazeh lagti hai.
Aik honest limit bhi hai. In mein se kuch bhi work ko apne aap behtar ya mehfooz nahin banata. Kamzor instruction ab apne aap chalti hai aur darmiyan mein pakarne wala koi nahin. Code, browser aur live connectors ko always-on surface mein rakhne se bad run ki reach barh jati hai. ChatGPT Work ke launch week mein security reviewers ne yehi point uthaya. Nai power sirf leverage hai. Woh madad karegi ya nuqsan, yeh brief, review aur judgment par hai. Is liye yeh course har nai power ke paas aik discipline rakhta hai, sirf power nahin bechta.
Agent power web par pehle se thi, lekin alag tools aur modes mein bikhri thi. Ab woh roz use hone wale web address par, everyday work ki aik jagah mein hai aur aap ke jane ke baad bhi chalti hai. Yeh faida hai. Catch yeh hai ke work company ke servers par jama hota hai. Yeh aap ke liye convenient aur un ke liye sticky hai. Convenience use karein, lekin finished cheez apni storage mein bhejein. Aur yaad rakhein: yeh surface work ko aap ke baghair chalata hai, work ko acha nahin banata. Woh hissa ab bhi aap ka hai.
Zehni tabdeeli, aik tasveer mein
Facts checked: 21 July 2026. Web par Cowork aur ChatGPT Work dono chand hafton purane hain. Rollouts plan ke mutabiq staged hain, usage metered hai aur names badalte hain. Yeh kitab ka sab se tez aging page hai. Kisi limit, plan tier ya feature par depend karne se pehle live product pages check karein.
Yeh course kya cover karta hai
| Part | Topic | Aap kya seekhte hain |
|---|---|---|
| 1 | Tabdeeli | July 2026 mein kya badla, remote session kya hai aur do vendors aik shape tak kyun pohanche |
| 2 | Surface | Sessions, account spine, files ke teen tiers, folders ke baghair connectors aur pocket mein gate |
| 3 | Baghair Nigraani Kaam | Delegation loop, woh work brief karna jisay aap nahin dekhenge aur kisi device ke online hue baghair scheduled tasks |
| 4 | Chunao aur Open Path | Har task ke liye web ya desktop, vendor cloud ke baghair open-source path aur course ka handoff |
Is course ko parhne ke do tareeqe. Pehli baar neeche Yeh kyun hua parhein, phir Parts 1-3 order mein aur Projects 1-3 karein. Taqreeban 90 minutes parhna hai aur kuch install kiye baghair agent user banata hai. Doosri reading real sessions ka aik hafta milne ke baad karein: Part 4, deeper notes aur Projects 4-6.
Kya yaad rakhein aur kya lookup karein. Is course mein do layers saath chalti hain aur un ki aging speed alag hai. Pehli yaad rakhein, doosri lookup karein. Lasting layer: chat box intezar karta hai aur hamesha karega; remote session account mein rehta hai, tab mein nahin; har deliverable teen tiers mein se aik mein utarta hai aur finished work ko bahar jana chahiye; human gate wahan pohanchna chahiye jahan aap waqai dekhte hain; har vendor ka agent product alag names mein wohi shape hai. Mechanical layer: neeche ke plan tiers, run limits, button labels aur product names. Yeh chand hafton purane hain aur har mahine badalte hain. Inhein live docs ke pointers samjhein, facts to memorize nahin.
Aik boundary pehle. Yeh course surface aur delegation loop sikhata hai. Runtime decisions nahin: verified unattended work lambe arse tak kahan rehna chahiye, move karne par kya saath jata hai aur trust dobara kaise milta hai. Section ke aakhir mein Leaving the Laptop yehi sikhata hai. Yeh course front door hai. Darwaze ke peeche ki kisi cheez par depend karne se pehle woh course parhein.
Part 1: Tabdeeli
1. Aik hi address, do mukhtalif cheezein
Aaj claude.ai ya chatgpt.com kholein, to chat box aur agent aik hi address par milte hain. ChatGPT mein yeh farq screen par hi nazar aata hai: Chat mode aur Work mode, aik hi app mein saath saath. (Desktop app mein coders ke liye aik alag Codex view bhi hai. Web aur phone par aap ko zyada tar Chat aur Work nazar aate hain.) Dono aik cheez nahin hain, aur inhein alag pehchanna is course ki pehli skill hai.
Chat box woh guftagu hai jis se aap Foundations mein waqif hue. Yeh har turn par aap ka intezar karta hai. Aap type karte hain, yeh jawab deta hai, phir dobara intezar karta hai. Yeh schedule par khud shuru nahin ho sakta. Yeh kisi event par react nahin kar sakta. Tab band karne ke baad yeh jari nahin reh sakta. In mein se koi cheez missing feature nahin: guftagu ki fitrat hi yeh hai. Chat box kabhi aap ke baghair kaam nahin chalayega, aur isay chalana bhi nahin chahiye. Yeh woh jagah hai jahan aap sochte, draft banate aur design karte hain.
Agent surface is ke paas hi hota hai. Aap isay message nahin, assignment dete hain. Yeh plan banata, tools istemal karta, steps par kaam karta aur deliverable paida karta hai. Yeh sirf wahan rukta hai jahan faislay ke liye aap ki zaroorat ho. Aur jab aap dekhna band kar dein tab bhi yeh kaam karta rehta hai.
Yeh aik saada test hai, aur apne baqi career mein isay yaad rakhna faidemand hai, kyun ke har company apni marketing mein is lakeer ko dhundla karegi:
Agar mein type karna band kar doon, to kya kaam ruk jata hai?
Chat box: haan. Agent surface: nahin. Yeh aik sawal har address par, har company ke liye, is saal aur baad bhi dono ko alag karta hai.
Jab aap mazeed samajhna chahein to isi khayal ka gehra version yeh hai: saadah chat turn synchronous hota hai. Jab tak aap intezar karte hain yeh chalta hai, phir ruk kar aap ke aglay turn ka intezar karta hai. Agent run delegated hota hai. Aik baar shuru karne ke baad yeh outcome ki taraf barhta rehta hai. Yeh baad mein chal sakta hai. Yeh kisi tabdeeli ko dekh sakta hai. Jari rehne ke liye isay aap ke aglay turn ki zaroorat nahin. Asal lakeer yehi hai, aur buttons aur menus mein mustaqbil ki tabdeeliyon ke baad bhi rahegi. "Type karna band" test isay rozmarrah mein mehsoos karne ka sirf saada tareeqa hai.
Aik ehtiyat, taa ke test aap ko ghalat raasta na dikhaye. Aksar aap delegated run ko chat box se, seedhay alfaaz mein shuru karte hain. "Har Monday, in emails ka khulasa karein." Alfaaz chat mein jate hain. Lekin jo cheez aap set up karte hain woh delegated task hai, mamooli reply nahin. Is liye sirf yeh na dekhein ke aap ne kahan type kiya, yeh dekhein ke kaam kya karta hai.
July 2026 mein aik aur cheez nahin badli, aur woh aham hai: chat box ka kaam apni jagah raha. Task design karna, brief ka draft banana, kisi khayal ko haath se test karna: yeh chat ka kaam hai, aur is course ka har assignment yahin se shuru hota hai. Agent surface ne chat box ki jagah nahin li. Woh bagal mein aa gaya.
Aik window, do kamray. Pehlay kamray mein jab tak aap bolein nahin, kuch nahin hota. Doosray kamray mein aap ke jane ke baad bhi worker kaam karta rehta hai. Yeh course doosray kamray ke baray mein hai. Jo cheez aap hand over karenge usay tayyar karne ke liye pehla kamra istemal karte rahenge.
Khud ko check karein
Aik colleague kehta hai: "Mein pehle hi roz claude.ai istemal karta hoon, is liye mein pehle hi agents istemal karta hoon." Do jumlon mein us ki islaah karein.
Jawab dikhayein
Woh chat box istemal karta hai: aisa kaam jo us ke type karna band karte hi ruk jata hai, jis ka matlab hai ke woh ab bhi har turn par kaam khud shuru kar raha hai. Agent surface is ke paas wala product hai: isay kaam assign karein, aur yeh plan banata, action leta aur tab band hone ke baad bhi jari rehta hai. Rozana chat istemal karna haqeeqi skill hai. Yeh agent istemal karna nahin.
2. Remote session: tab aik window hai, runtime nahin
Yeh course ka bunyadi architecture concept hai, aur is ki har doosri cheez isi aik haqeeqat se nikalti hai.
Agent do jagahon par reh sakta hai. Yeh aap ki machine par reh sakta hai: desktop apps aur terminals jin se aap aglay teen courses mein milenge. Wahan app hi runtime hoti hai (woh cheez jo asal mein kaam chalati hai). Laptop band karein to kaam usi jagah pause ho jata hai, kyun ke kaam karne wali machine so gayi. Ya yeh vendor ke servers par reh sakta hai: yeh course isi ke baray mein hai. Wahan aap ka browser tab session ki taraf aik window hai, kaam chalane wali machine nahin.
Nateeje bayan karna saada hai, aur har aik aap ka kaam karne ka tareeqa badalta hai:
- Tab band karein: kaam jari rehta hai. Session kabhi tab mein tha hi nahin. Yeh aap ke account mein aise infrastructure par chal raha hai jo aap ke sone par nahin sota.
- Phone se session kholein: yeh wohi session hai. Na copy, na summary. Wohi chalta hua kaam, sirf choti window se dekha ja raha hai.
- Scheduled task us tab se chalta hai jo aap ne pichlay haftay band kiya tha. Schedule bhi vendor ke servers par rehta hai. Aap ke devices critical path mein nahin.
claude.ai par usi browser tab mein Cowork session shuru karein jahan aap chat karte hain. Yeh Anthropic ke servers par remote session ki shakal mein chalta hai. Aap ke sessions aur files aap ke account mein save hote hain: aisa tasalsul jisay aap ko khud set up nahin karna para. Jab kaam aise faislay tak pohanchta hai jo sirf aap kar sakte hain, approval aap ke phone par Claude app mein aata hai.
Yeh course likhe jane ke waqt web par Cowork beta mein hai: pehle Max plan, phir mazeed plans, aur usage metered hai. Yahan mechanical layer tezi se badalti hai. Is par bharosa karne se pehle live product page par apne plan ki access confirm karein.
chatgpt.com par ChatGPT mein Work mode par switch karein (9 July 2026 ko release hua, GPT-5.6 se powered). Yeh alag app nahin: aik hi interface mein Chat mode ke paas rehta hai, aur multi-step project par ghanton tikay rehne, goal ko steps mein torne aur kuch hisson ko khud mukammal karne ke liye bana hai. Sessions web, desktop aur mobile ke darmiyan cloud-sync rehte hain: laptop par shuru karein, phone se check karein, wohi session. Aur Scheduled Tasks cloud se chalte hain chahe aap ka koi device online ho ya nahin.
Access plan ke mutabiq marhala war hai. Web aur mobile par Work eligible paid plans mein rollout hua, Free ya Go mein nahin. (ChatGPT desktop app par Chat, Work aur Codex har plan mein dastiyab hain, Free samet. Codex us desktop app ka alag view hai: web ya mobile par select nahin hota.) Yeh un paid plans mein kisi izafi list price ke baghair bundled hai, aur usage har plan ke agent-usage allowance se kam hoti hai, is liye yahan "metered" ka matlab plan ka included allowance hai, alag bill nahin. Availability badalti rehti hai, aur plan ki yeh details is kitab ke sab se jaldi purane hone walay claims hain. Is par bharosa karne se pehle OpenAI ke live product page par apni mojoodah access confirm karein.
Aik imaandaar had bhi samajh lein taa ke yeh concept khud ko zyada na beche. Remote session company ki machines par chalta hai. Apne bal par yeh un cheezon tak pohanchta hai jahan woh machines pohanch sakti hain: aap ke connectors, is ka apna task filesystem, aap ki platform files. Apne bal par yeh aap ki local hard drive, desktop apps ya us browser profile tak nahin pohanchta jahan aap apne bank mein logged in hain. Aise kaam ko local bridge ya local agent chahiye, aur aglay courses yehi sikhate hain. (Baad ke liye aik nuance: desktop app bridge ka kaam kar sakti hai. Jab tak woh khuli ho, remote session ko local files ya browser tak mahdood access de sakti hai. App band karein to session chalta rehta hai magar local rasai kho deta hai. Abhi aap ko is ki zaroorat nahin. Sirf itna jaan lein ke deewar mukammal nahin.) Part 4 aap ko poori decision table deta hai.
Desktop agent aap ke ghar ke andar worker hai: jab ghar ki bijli jaye to worker ruk jata hai. Remote session vendor ke office mein worker hai: aap ke ghar ki bijli se farq nahin parta, aur aap ka laptop aur phone sirf windows hain jin se aap usay check karte hain.
Mazeed gehrai: aap is khayal se dobara, do baar milenge
Section mein baad ke liye is concept ko zehan mein halka sa rakhein. Loop Engineering aap ko sikhayega ke unattended work ko runs ke darmiyan kisi aisi cheez ki zaroorat hoti hai jo jagti rahe, aur poochay ga ke woh cheez kahan rehti hai. Leaving the Laptop aap ke desk ki machine ko "aakhri single point of failure" kahega aur usay hatane ka poora faisla sikhayega. Is concept mein aap dono khayalat ka consumer-grade version dekh rahe hain: vendor ke servers jagte rehte hain taa ke aap ke servers ko na jagna pare. Wohi khayal. Koi config files nahin.
Khud ko check karein
Lahore mein Ayesha shaam 5 bajay web session mein client report shuru karti hai. 6 bajay load-shedding se us ki bijli chali jati hai. 8 bajay woh phone se session kholti hai. Usay kya milta hai, aur us ki bijli kabhi masla kyun nahin thi?
Jawab dikhayein
Usay kaam mukammal milta hai, ya phone se clear karne ke liye approval par intezar karta hua. Session poora waqt vendor ke servers par chala: us ka laptop sirf is ki taraf window tha, aur band window doosri taraf chalne wali cheez ko nahin rokti. Desktop agent, jis se woh Cowork desktop course mein milegi, machine ki bijli jate hi pause ho jata, kyun ke wahan machine hi runtime hoti hai.
3. Do vendors, aik shape
Ab woh concept jis ke liye yeh course mojood hai. July 2026 mein Anthropic aur OpenAI ne chand din ke farq se apni agent surfaces browser mein release ki. Dono products aise parts ka set dikhate hain jo saaf tor par milta julta hai. Aur dono aik hi anatomy ke saath aaye.
Yeh kitab us anatomy ko chhay parts ke tor par sikhati hai, aur baad ke courses har aik ko tafseel se banate hain. Aaj in parts ko reading lens ke tor par istemal karne ke liye aap ko un courses ki zaroorat nahin. Har part aik jumlay mein yahan hai, aur ChatGPT Work mein woh kahan nazar aata hai; neeche ka har naam OpenAI ka apna hai, jo July 2026 ke launch se liya gaya hai:
| ChatGPT Work isay kehta hai | Yeh kitab isay kehti hai | Aik jumlay mein matlab |
|---|---|---|
| Scheduled Tasks | Heartbeat | Aap ke baghair kaam kya shuru karta hai: aik baar, schedule par, event par, ya tabdeeli ko monitor karte hue. |
| Plugin Directory (connected apps) | Connectors | Aap ki haqeeqi services, jaise Slack, Gmail, Drive aur darjanon doosri, tak permission-scoped rasai taa ke agent sirf mashwara na de balkeh action le sake. |
| Outcome-based task execution | Run-until-done loop | Isay outcome dein, aur yeh steps mein us outcome ki taraf kaam karta hai: aik jawab nahin, finish line. |
| Cloud-synced sessions & Projects | State spine | Aisi memory jo devices aur runs ke darmiyan qaim rehti hai, taa ke kaam dobara shuru hone ke bajaye jari rahe. |
| Approval prompts (mobile) | Human gate | Risky ya judgment-heavy faislay insaan ke paas, phone par, aate hain aur us ke baad hi kaam jari rehta hai. |
| Work's cloud execution environment and artifact tools | Body | Woh jagah jahan run ka kaam asal mein hota hai aur finished artifacts bante hain: docs, sheets, decks, dashboards aur choti web apps. |
Web par Cowork ko isi lens se dekhein to Anthropic ke names ke saath wohi table samne aati hai: scheduled tasks, connectors, account mein save sessions, phone par approval. Do vendors. Aik shape.
Agent career ke teen hafton baad aap ke liye yeh kyun aham hai? Kyun ke products release hote rahenge aur names badalte rahenge, aur aap har baar zero se dobara nahin seekh sakte. Jab mukhtalif builders baar baar wohi parts tak pohanchte hain, to parts discipline hote hain aur products packaging. Shape aik baar seekhein, aur har naya agent product naye field ke bajaye aadhe ghantay ki reading ban jata hai. Yeh is kitab ka markazi dawa hai, aur July 2026 ab tak is ka sab se mazboot saboot hai: skill tool nahin, shape hai.
Do imaandaar notes, taa ke lens sirf lens rahe. Pehla, aik jaisi shape ka matlab aik jaisa behavior nahin. Rollout order, plan tiers, limits aur bareek details dono mein mukhtalif hain aur har haftay badalti hain. Mapping aap ko batati hai ke kya talash karna hai. Yeh wada nahin karti ke woh aik hi tareeqay se set up milega. Doosra, chhay parts yahan sirf vocabulary ke tor par pesh kiye gaye hain. Loop Engineering aap ko har aik design karna sikhata hai, aur us se pehle aap ko koi sanjeeda cheez schedule nahin karni chahiye (Concept 9 mazeed batata hai).
Aik aur note jo asal mein Concept 6 ka preview hai. ChatGPT Work ke launch ne code execution, browser access aur enterprise connectors ko aik always-on space mein rakha, is liye aik kharab run zyada door tak pohanch sakta tha, jis baat ko security reviewers ne foran flag kiya. Yeh product par tanqeed nahin. Yeh shape ki qeemat hai, aur har vendor yeh qeemat deta hai. Isi liye connectors ka apna concept aage aata hai, aur isi liye approval gate optional nahin.
Mukhtalif companies ki cars alag dikhti hain, magar har car mein engine, wheels, brakes aur steering wheel hota hai. Parts jaan lene ke baad aap thori der dekh kar koi bhi car chala sakte hain. Agent products bhi aise hi hain. Yeh table parts ki list hai. Yeh kitab aap ko parts sikhati hai, taa ke kisi vendor ki packaging dobara aap ko uljha na sake. Aur dekhein: do companies miltay jultay engine ke gird behtar car banane ki race mein hain. Woh race "harness war" hai. Web par aap car chunne walay driver hain, car banane walay engineer nahin. Dono kaam aham hain. Bas dono mukhtalif kaam hain.
Shape aik hai, to dono asal mein kahan mukhtalif hain? Chhay parts shared hain. Packaging shared nahin. Yeh imaandaar side-by-side muqabla hai, jo har vendor ke July 2026 launch se banaya gaya hai. Har row ko mid-July 2026 samjhein aur live pages se verify karein; yeh kitab ki sab se tez aging table hai.
| Web par Cowork (Anthropic) | ChatGPT Work (OpenAI) | |
|---|---|---|
| Yeh kya hai | Remote session, claude.ai par Cowork surface se dastiyab | ChatGPT ke andar aik mode, chatgpt.com par Chat mode ke paas |
| Kahan chalta hai | Anthropic ke servers | OpenAI ke servers |
| Model | Claude (har task ke liye Opus/Sonnet tier) | GPT-5.6 family |
| Kaisay becha jata hai | Web beta, plan ke mutabiq marhala war; metered | Web aur mobile par eligible paid plans, naye desktop app par har plan mein bundled; har plan ke agent allowance se usage kam hoti hai |
| Connectors | MCP par Anthropic ka connector catalog | Plugin Directory (launch par App Directory se naam badla), yeh bhi MCP-based hai |
| Run-until-done | Brief mein saaf stopping condition | Outcome dein, aur yeh steps mein us ki taraf kaam karta hai |
| Persistent workspace | Projects + saved context file (Anthropic mein is ka naam CLAUDE.md hai) | Projects + memory (jisay review aur correct kiya ja sakta hai) |
| Finished artifacts | Docs, sheets, decks, files | Docs, sheets, decks, dashboards aur Sites (URL par publishable web apps) |
| Coding sibling | Claude Code (alag surface) | Codex (mode ke tor par usi app mein merge) |
| Mobile gate | Claude app mein approvals | ChatGPT app mein approvals |
Parts ko zehan mein rakh kar table parhein to aik pattern foran nazar aata hai: taqreeban har difference packaging column mein hai. Darmiyan ke parts shared hain. Model mukhtalif hai. Plan rules mukhtalif hain. Connector list ka naam mukhtalif hai. Aik URL-publishing feature deta hai jo doosra nahin deta. Kuch differences labels se gehre hain: task aap ki local files mein kahan tak pohanch sakta hai, approvals kaisay kaam karte hain, coding agent usi app mein rehta hai ya nahin. Yeh haqeeqi hain, aur Concept 10 inhein istemal karta hai. Magar heartbeat phir bhi heartbeat hai. Gate phir bhi gate hai. Spine dono mein phir bhi spine hai. Shared parts lasting layer hain. Baqi woh mechanical layer hai jisay is course ne yaad karne ke bajaye lookup karne ko kaha. Chhay-part lens ka nuqta yehi hai, kaha nahin balkeh dikhaya gaya hai. Agar aap ne sirf aik product seekha hota, to doosra phir bhi choti si sair hota, kyun ke kaam sambhalne walay parts wohi hain jo dono share karte hain.
Jo kuch aap dekh rahe hain us ka aik naam. Ghor karein ke dono products kahan muqabla karte hain. Sirf model par nahin. Models ab bhi aham hain, aur dono mazboot hain. Magar products model ke gird har cheez par muqabla karte hain: scheduler, connector layer, session spine, phone gate, files bahar nikalne ka tareeqa. Yeh kitab us wrapper ko harness kehti hai, aur Harness Engineering isi par course hai. Pichlay do saalon ke zyada tar hissay mein harness woh cheez thi jo engineers terminal mein code se khud banate thay. July 2026 ke yeh launches dikhate hain ke harness finished product ke tor par web par aa gaya hai, aur do companies is baat par muqabla kar rahi hain ke kis ka harness behtar hai. Yeh muqabla aap ke liye acha hai. Isi liye surface itni tezi se behtar hui.
Magar aik lakeer saaf rakhein taa ke isay durust parhein. Web par companies harness war larti hain. Aap nahin larte. Yahan aap harness nahin bana rahe. Aap do pehle se banay harnesses mein se chun rahe hain. Loop, tools, context aur checks ko wire karna engineer ka kaam hai, baad ka Harness Engineering course isi ka mazmoon hai, aur yeh haqeeqi, gehra kaam hai. Is surface par aap ka kaam chota aur mukhtalif hai: task ke liye munasib pre-built harness chunein, phir usay achi hidayat dein. War ka hona jaan lena chunao mein madad deta hai. Is se aap harness engineer nahin bante. Woh aik alag course aur alag chair hai.
Har product kis ke liye hai? Parts shared hain, magar vendors ka nishana aise log hain jo kuch had tak milte aur kuch had tak mukhtalif hain, aur nishana jaan lena chunao mein madad deta hai. Is ki do layers hain:
Companies ne kisay nishana banaya. Dono products usi core user ko pesh kiye gaye hain: knowledge worker. Yeh woh shakhs hai jis ka din documents, spreadsheets, email aur chat ke darmiyan information le jate guzarta hai, aur jo code nahin karta. "Finished spreadsheet, deck ya report paida karein" isi shakhs ke liye likha gaya hai. Magar us user ke gird shape mukhtalif hai. ChatGPT Work aik aise app ke andar mode hai jis mein coding agent (Codex) bhi hai. Is liye OpenAI aise workplace ko nishana bana raha hai jahan aik hi shakhs ya team apps badlay baghair knowledge work aur code dono kar sakte hain. Aur yeh mojoodah paid plans ke saath bina nayi list price ke aata hai, is liye nishana wasee hai. Web par Cowork Max-first aur metered launch hua. Yeh chota, zyada iraday wala ibtidai group hai. Is ka coding agent (Claude Code) alag surface rehta hai. Yeh do qisam ke users ke darmiyan zyada saaf taqseem hai.
Kis ko kya chunna chahiye. Yeh is kitab ki rehnumai hai, vendor ka dawa nahin. Zyada tar document-and-connector knowledge work ke liye dono theek hain. Imaandaar tie-breakers mamooli hain: aap kis company ka model pasand karte hain, aap ke connectors pehle kis ke peeche rehte hain, aur aap kis plan ki payment pehle hi karte hain. Faisla bunyadi tor par aik surat mein zyada saaf hota hai. Agar aap knowledge work aur code dono karte hain aur unhein aik chhat ke neeche chahte hain, to ChatGPT ka desktop app Chat, Work aur Codex ko saath rakhta hai, jo halki setup hai. Agar aap work agent aur coding agent ko alag, alag governance wali surfaces rakhna chahein, to Cowork aur Claude Code ki taqseem saaf fit hai. Dono suraton mein yeh zyada tar packaging ka chunao hai, capability ka nahin, isi liye yeh shape ka footnote hai, headline nahin.
Yeh course dono ke liye aik audience ko kharij karta hai, aur yeh aham hai: regulated data sambhalne wala koi bhi shakhs. PHI (protected health information), privileged matter, regulated financial records, kisi web surface ke target user nahin jab tak compliance likh kar ijazat na de, kyun ke working files aur platform storage ki custody bunyadi tor par vendor ke paas hoti hai. Concept 10 isay poora samjhata hai. Yahan sirf yaad rakhein ke "yeh kis ke liye hai" ki aik sakht boundary hai jo kisi vendor ki marketing aap ke liye nahin kheenchti.
Packaging se aage aik haqeeqi farq flag karna zaroori hai, kyun ke yeh aap ke aik chunao ko badalta hai. OpenAI ka coding agent, Codex, Work ke saath usi ChatGPT app mein rehta hai, agarche desktop app par alag view hai, aur web ya phone par select nahin hota. Anthropic apne coding agent, Claude Code, ko mukammal alag surface rakhta hai. Knowledge worker ke liye yeh nazar nahin aata. Jo shakhs knowledge work aur code dono karta hai us ke liye yeh aham hai. Desktop par ChatGPT aik chhat ke neeche Chat, Work aur Codex rakhta hai. Claude ki taraf aap ko Cowork aur Claude Code ke darmiyan jana parta hai. Koi aik behtar nahin. Yeh haqeeqi shape difference hai, aur Concept 10 ki "kaam kis cheez ko choota hai us se chunein" table mein is ka faida milta hai.
Dono products aik hi tarah banay gaye hain aur zyada tar apne labels, model aur har plan ke bechne ke tareeqay mein mukhtalif hain. Aik asal structural farq: OpenAI ne coding tool ko work agent ke saath usi app mein rakha. Anthropic dono ko alag rakhta hai. Document work mein aap ko farq mehsoos nahin hoga. Agar aap code bhi karte hain to yeh jaanne laiq hai.
Khud ko check karein
Agli quarter mein teesra vendor "Flows" naam ka agent product release karta hai. Is ka launch page yeh list deta hai: triggers, integrations, autopilot mode aur workspace memory. In ki docs ka aik lafz parhne se pehle chaaron names ko kitab ke parts se map karein.
Jawab dikhayein
Triggers: heartbeat. Integrations: connectors. Autopilot mode: run-until-done loop (aur docs ke liye aap ka pehla sawal hai: stopping condition kya hai, aur isay kaun check karta hai?). Workspace memory: state spine. Aap ne aik minute se kam mein aisa product parh liya jo pehle kabhi nahin dekha aur durust sawalat ke saath pohanche. Yeh exercise hi woh skill hai jo yeh concept sikhata hai.
Part 2: Surface
Tabdeeli peeche reh gayi. Ab surface ke apne working parts: aap ke sessions kahan rehte hain, files asal mein kahan utarti hain, jab koi na dekh raha ho to connector kya karta hai, aur woh aik faisla jisay aap ki zaroorat hai aap tak kaisay pohanchta hai. Chaar concepts, aur un mein se aik, teen file tiers, is course ka sab se aham discipline hai. Wahan pohanchein to aahista parhein.
4. Account spine: sessions aur files jinhein aap ko banana nahin para
Aap ke pehlay Foundations workflow mein standalone conversation aksar zero se shuru hoti thi. Aap context dobara paste karte, project phir samjhate, file dobara upload karte. Aksar continuity haath se sambhalna aap ka kaam tha.
Agent surface continuity ko feature bana kar deti hai. Aap ke sessions account mein save hote hain. Session jo files banata hai woh account mein save hoti hain. Kisi bhi device se koi cheez kholein to wohi kaam usi nuqtay par milta hai. Yeh course isay account spine kehta hai (Concept 3 ki state spine ka is surface wala version): aisi memory jo sessions aur devices ke darmiyan qaim rehti hai, vendor banata hai, aur isay maintain karne ke liye aap ko kisi ko hire nahin karna parta.
Yeh haqeeqi tohfa hai aur aap ko istemal karna chahiye. Is ki qeemat bhi hai, aur kaam is par rakhne se pehle qeemat jaan leni chahiye. Spine vendor ki custody aur vendor ke format mein rehti hai. Aap ke sessions, files aur kaam ki history un ki machines, un ke account system aur un ki tai ki hui shapes mein hoti hai. Is mein koi khaufnaak baat nahin. Magar har cheez aik dependency hai: aap ke paid plan, product ke qaim rehne aur vendor ke rules ke aap ke rules se compatible rehne par. Concept 5 woh discipline deta hai jo dependency ko mehfooz rakhta hai. Filhal dono hisay yaad rakhein: spine muft continuity hai, kisi aur ke ghar mein.
Har Cowork session aap ke claude.ai account mein save hota hai aur us ki banai files saath attached hoti hain. Kal kisi bhi device se session par wapas aayein to context, plan aur files wahin milenge jahan chore thay. Is surface par sessions list asal mein aap ka work log hai.
ChatGPT Work ke sessions web, desktop aur mobile mein cloud-sync hote hain: laptop par shuru karein, phone par dobara kholein, aur wohi session, state aur files milti hain. Do features us spine ko barhate hain. Projects mutaliq chats, files aur instructions ko aik persistent space mein jama karte hain jo sessions mein context share karta hai, aur recurring work ka qudrati ghar hai (Concept 6 ka persistent-workspace khayal, built in). Aur jahan memory enabled ho, yeh pichli chats ka relevant context saath la sakti hai, reviewable summary aur aise sources ke saath jinhein aap inspect ya correct kar sakte hain. Session list, projects aur memory spine hain. Devices windows hain. Cowork tab ne custody ke baray mein jo kaha woh yahan bilkul waisay hi lagu hota hai: yeh spine OpenAI ke servers par, OpenAI ke format mein rehti hai.
Teen choti aadatein spine ko aap ke khilaf nahin, aap ke liye kaam karwati hain. Sessions ko chats nahin, work products jaisay names dein. "Acme renewal brief. May invoice run" teen mahine baad khud mil jata hai. "Quick question" kabhi nahin milta. Har session mein aik workstream. Client report aur chutti ki planning aik session mein milane se context dono mein behta hai, aur Foundations mein aap ne seekha ke stale context quality ke saath kya karta hai. Jo mar chuka ho usay hata dein. Lamba session apni poori history har turn mein lata hai, aur is surface par context ki qeemat bilkul chat jaisi hai: maqsad ke saath chuna kam context, aadat se ghaseeta zyada context se behtar hai.
Chat tab mein aap har shaam papers ghar le jate aur har subah wapas late thay. Agent surface par aap ke paas drawers wala desk hai: har cheez wahin rehti hai jahan aap ne chori, aur har jagah se dastiyab hai. Desk bohat kaam ka hai. Bas yaad rakhein ke desk vendor ki building mein hai, aap ki nahin.
Khud ko check karein
Account spine ka tohfa aur qeemat, har aik ko aik jumlay mein batayein.
Jawab dikhayein
Tohfa: woh continuity jo aap ko banani nahin pari, sessions aur files jo tab, din aur device ke baad bhi qaim rehte hain. Qeemat: yeh continuity vendor ki custody aur format mein rehti hai, is liye jo cheez sirf wahan mojood ho woh aap ke plan, un ke product aur un ke rules par depend karti hai. Agla concept is qeemat ko working discipline banata hai.
5. Teen file tiers: deliverable asal mein kahan rehta hai?
Yeh is course ka signature concept hai. Agent jis file ko choota hai usay teen tiers mein se aik mein rehta samjhein. Yeh model hai, physics ka qanoon nahin, magar amal mein mazboot rehta hai aur dhundlay sawal ko tez bana deta hai. Har deliverable ke liye aap aik faisla karte hain: yeh kis tier mein utarta hai?
Tier 1: task filesystem. Remote session kaam karte waqt vendor ke servers par temporary working space istemal karta hai: drafts, darmiyani calculations aur aadhi bani files ke liye scratch area. Kaam mukammal hone par yeh clean ho jata hai, aur is ke exact retention rules vendor ke mutabiq mukhtalif hain, is liye isay kabhi storage na samjhein. Tier 1 apni fitrat ke baray mein imaandaar hai. Kal chahiye honay wali kisi cheez ka yeh kabhi ghar nahin.
Tier 2: platform storage. Vendor ke platform par aap ke account mein permanently save files. Yeh task, tab aur hafta guzarne ke baad bhi rehti hain. Yeh Concept 4 ki account spine hai jo files par lagu hoti hai, aur qudrati default hai: jab tak aap kaam kahin bhejein nahin woh yahin utarta hai. Tier 2 mehfooz hai. Tier 2 vendor ki custody, vendor ke format aur vendor ke login ke peeche bhi hai.
Tier 3: exit. File platform se nikal kar aisi jagah utarti hai jisay aap control karte hain. Connector se Drive mein save. Email ya Slack post. Apni machine par direct download. Local folder mein write. Repo mein commit. Darwazay kai hain aur darwaza aham nahin. Manzil aham hai: Tier 3 woh wahid tier hai jahan deliverable aap ke system of record mein dakhil hota hai, us jagah jahan aap ki firm, client ya mustaqbil ka aap asal mein usay talash karega.
Discipline aik line mein, aur poore course se yehi line yaad rakhni hai:
Finished work platform se nikalta hai. Har doosri cheez reh sakti hai.
Tiers ko risk ladder ki tarah parhein to line khud samajh aati hai. Sirf tier 1 mein mojood deliverable pehle hi gum hai: wipe bug nahin, tier ki definition hai. Sirf tier 2 mein mojood deliverable aaj mehfooz aur kal hostage hai: mehfooz kyun ke qaim rehta hai, hostage kyun ke sirf wahin, kisi aur ki custody aur login ke peeche qaim rehta hai. Tier 3 ka deliverable aap ka hai: aap ke Drive ki version history, firm ke matter folder, repo mein, aap ke retention rules ke tehat, colleagues ko milne laiq, aur vendor ka product agli quarter shape badlay tab bhi mojood.
Cowork web session temporary task filesystem (tier 1) mein kaam karta, outputs session ke saath account mein save karta (tier 2), aur connected services se files save ya send kar sakta hai. Drive, Gmail, Slack (tier 3). Tier-3 ka raasta aap ke diye connectors se guzarta hai, isi liye agla concept connector scope ko nayi sanjeedgi se leta hai.
Wohi teen tiers. Tier 1 isolated working environment hai jisay Work task chalte waqt istemal karta hai, apne file system ke saath, sirf is job ke liye banaya gaya. Is ki exact implementation OpenAI manage karta hai aur aap ke control walay interface ka hissa nahin. Tier 2 is ki banai finished files hain: docs, sheets, decks, dashboards aur Sites, jo cloud-synced session aur Projects ke saath rehti hain. Tier 3 exit hai: finished file platform se nikal kar aap ke control ki jagah mein jati hai. Is browser surface par bunyadi automated exit Plugin Directory se connector save hai, Drive, Gmail ya Slack ki taraf. Download manual exit hai, aur jahan desktop bridge connected ho wahan local write izafa kar sakti hai. Yahan OpenAI ke haq mein aik note: Sites finished web app ya dashboard ko shareable URL par publish kar sakta hai. Yeh waqai mufeed hai. Magar isay Concept 5 se parhein. Platform par hosted Site ab bhi public door ke saath tier 2 hai, aap ki custody mein tier 3 nahin. Exact keep-and-export rules chand hafton purane hain, is liye live docs se verify karein. Neechay ka three-tier structure nahin badlega, aur seekhne wali cheez wohi hai.
Aik worked example. Ayesha client ke liye monthly invoicing chalati hai. Draft numbers tier 1 mein rehte hain: agent ka scratch math aur temporary CSV. Woh wahin jeete aur marte hain, aur yeh durust hai. Koi scratch paper audit nahin karta. Invoice template woh aglay mahinay dobara istemal karegi, is liye usay tier 2 mein save karti hai: platform storage, aise session ke saath attached jisay woh tees din baad dobara kholegi. Finished invoice PDF do baar tier 3 mein jati hai: firm ke Drive folder mein save, aur mail connector se client ko email. Do baar kyun, aur tier 3 mein kyun? Kyun ke aik din koi poochay ga, "March ka invoice kahan hai?" Auditor, partner ya November mein khud Ayesha. Jawab firm ke records hona chahiye, "mere Cowork account mein kahin" nahin. System of record firm ka hai. Sirf Tier 3 wahan pohanchta hai.
Har deliverable banane walay brief ko aik instruction par khatam karein, aur tier ka faisla khud sambhal jata hai:
End by listing every file you created and where each one landed:
temporary working space, platform storage, or a system I control (connector save, download, local write, or repo commit).
Yeh aik line chhupe default ko nazar aane wala faisla banati hai, aur Project 2 mein aap isay real work par chalayenge.
Paper teen jagahon par ho sakta hai. Worker ke scratch pad par: din ke aakhir mein phenk diya jata hai. Worker ke apne drawer mein: mehfooz, magar us ke desk mein, aap ke nahin. Aap ki filing cabinet mein: aap ka, jahan aap ke rules lagu aur log isay dhoond sakte hain. Rule saada hai: har finished cheez aap ki filing cabinet mein jati hai. Drafts pad ya drawer mein reh sakte hain.
Mazeed gehrai: is discipline ka mustaqbil hai
Do courses baad yeh concept bara ho kar aap ko wapas milega. Loop Engineering state spine sikhayega: runs ke darmiyan loop ki rakhi memory, aur us memory ka maqam bilkul tier ka faisla hai. Leaving the Laptop suitcase test sikhayega: ghar badalne par kya saath jata hai, aur is ka rule, repo sach rakhta hai, engineering kapron mein tier 3 hai. Aur is course ka Project 6, exit drill, us course ki vanishing-home drill ka beginner version hai. Ab invoices aur briefs par tier ki aadat seekhein, aur wahan pohanchne par engineering versions bilkul wazeh lagenge.
Khud ko check karein
Aik lawyer web session mein comparison memo banata aur platform link client ke saath share karta hai. Wahan rukne ki do ghalatiyan batayein.
Jawab dikhayein
Pehli, custody: memo sirf tier 2, vendor ke platform par hai, firm ke matter folder aur retention system se bahar, is liye engagement ke firm ke apne record mein suraakh hai. Doosri, access: link ki sharing aur permission platform ki hai, firm ki nahin, aur firm us cheez ko govern nahin kar sakti jo us ke paas nahin. Hal tier 3 hai: memo document management system mein save karein, phir firm ke rules ke tehat wahan se share karein. Platform link window thi. Client ko document ke bajaye window di gayi.
6. Web par connectors: rasai aur exit door aik saath
Aap Foundations mein connectors se milay: MCP standard ke zariye real service tak permission-scoped access. Drive, Gmail, Slack. Aap scopes dete hain. AI un ke andar kaam karta hai. Yahan us definition ki koi cheez nahin badalti. Jo badalta hai woh wazan hai jo connector tab uthata hai jab isay chat nahin, agent pakray hue ho.
Is surface par aap ke trust controls do levers tak aate hain: connector scope aur approvals. Desktop agents, jin se aap Cowork desktop course mein milenge, teesra lever shamil karte hain (agent kin local folders ko dekh sakta hai) aur is ke gird poora discipline banate hain. Browser-only session par woh lever mojood nahin, kyun ke is ke paas apne bal par scope karne ke liye local filesystem nahin. Us ka kaam abhi seekhay tier decision aur aap ke diye connector scopes mein taqseem hota hai.
Yahan se woh haqeeqat nikalti hai jisay yeh concept seedha bayan karta hai: web par connectors dohra wazan uthate hain. Connector aap ki services mein agent ki rasai hai, yeh hamesha sach tha. Browser-only session par yeh files ka bunyadi automated exit door bhi hai: Concept 5 ka tier-3 raasta zyada tar aap ke diye connectors se guzarta hai. Download manual exit hai, aur jahan supported ho desktop bridge local write shamil kar sakta hai, magar hands-off work ke liye connector darwaza hai. Service connect karte waqt jo permission screen aap ne dekhi, woh ab dono cheezon ko govern karti hai: agent kya parh sakta hai aur finished work kahan ja sakta hai. Is liye Foundations ki aadat good practice se rule ban jati hai: connect par click karne se pehle scopes parhein, aur kaam ke liye sab se chota scope dein.
Pehle real task se pehle do baatein jaanne laiq hain:
- Read scope, send scope nahin. Mail connector ko read access dene se agent threads summarize kar sakta hai. Write ya send access alag aur zyada bari grant hai. Aap ke address se bheja message wapas nahin bulaya ja sakta, aur log us par action lenge. Yahan safe default assume na karein. Connectors bohat mukhtalif hote hain: kuch read-only, kuch sirf drafts bana sakte aur kuch app, plan aur workspace settings ke mutabiq haqeeqi write ya send actions le sakte hain. Is liye jo connector enable karein us ke asal actions aur permissions inspect karein. Phir nateeja rakhne wali har cheez ke liye approval lazmi rakhein, aur send kar sakne wali setup ko alag, zyada trust wala faisla samjhein.
- Untrusted content ko hamesha careful mode milta hai. Yeh us rule ka mukhtasar roop hai jo Cowork course poora sikhata hai. Kisi aur ka likha text chhupi instructions rakh sakta hai. Inbound email, vendor PDF, unknown webpage. Aap ko mamooli jumlay nazar aate hain. Agent ko commands lag sakte hain. Is chaal ko prompt injection kehte hain. Yeh theory nahin, haqeeqi attack hai. Is satah par difa saada hai. Jab task aise content ko chhooay jo aap ne nahin likha, ask-before-acting mode mein rahein. Plan mein aise steps talash karein jo aap ne nahin maange. Agar plan aisa connector action pesh kare jo aap ne request nahin kiya, approve na karein. Redirect karein, ya dobara shuru karein.
Grant dene ka discipline Foundations se nahin badla magar stakes barhne par dohrana zaroori hai: service tab connect karein jab koi khaas workflow isay mangay, pehle se andazay par nahin. Har connector rasai hai, aur ab darwaza bhi.
Connector woh chaabi hai jo aap worker ko dete hain. Is surface par wohi chaabi do cheezein kholti hai: woh kamray jahan worker information lene ja sakta hai, aur woh mail slot jis se finished documents building se nikalte hain. Kaam se zyada kholne wali chaabi convenience nahin. Yeh aap ka chuna risk hai. Har chaabi dene se pehle parhein ke woh kya kholti hai.
Khud ko check karein
Desktop par (agla course), trust levers folders, connectors aur approvals hain. Web par do hain. Teesray ko kya hua, aur us ka kaam kis ne sambhala?
Jawab dikhayein
Folder lever gaib hua kyun ke browser-only remote session ke paas apne bal par scope karne ke liye local filesystem nahin. (Desktop bridge jab desktop app khuli ho to scoped local access de sakta hai, magar woh desktop ka lever hai, tab ka nahin.) Is ka kaam do cheezon ko mila: tier decision (Concept 5), jo files kahan rehti hain govern karta hai, aur connector scope, jo agent ki rasai aur files ke exit ko govern karta hai. Yehi virasat is surface par connectors ko dohra wazan deti hai.
7. Aap ki jaib mein gate
Is kitab ka har agent product aik wada rakhta hai: jin faislon ko waqai insaan chahiye woh insaan tak aate hain. Yeh kitab isay human gate kehti hai. Yeh surface gate ko nahin badalti. Yeh badalti hai ke gate aap ko kahan dhoondta hai.
Desktop agent par aam tor par approval card aate waqt aap screen ke paas hote hain: ittefaq se gate aur aap ki aankhein aik jagah hain. Is surface par woh ittefaq khatam hai. Kaam lunch, meeting ya neend ke waqt chalta hai, aur surface ka maqsad yehi hai, is liye approval tab par intezar nahin kar sakta. Woh phone par aata hai: notification, approval card, aisa faisla jisay aap kahin se tees seconds mein parh aur clear kar sakte hain.
Phone ka kirdar bilkul durust samjhein, kyun ke dono ghalatiyan aam hain. Phone gate hai, workbench nahin. Phone se plan review, step approve, aik jumlay mein redirect ya task stop karein. Wahan brief ka draft nahin banate, aur chalis-page deliverable review nahin karte. Design ka kaam bari screen ka hai. Phone is liye hai ke session das-second ke faislay ke liye chhay ghantay na ruka rahe.
Ab woh ehtiyat jis ke liye yeh concept mojood hai. Yeh aagay parhay jane walay course se li gayi hai kyun ke intezar ke liye bohat aham hai: gate nahin hila. Doorbell ko hilna para. Jab aap screen par kaam karte thay, escalation khud aankhon ke samne aati thi. Ab kaam wahan chalta hai jahan aankhein nahin. Is liye approval baghair nazar aaye para reh sakta hai. Aur unseen escalation delayed decision hai, jo kabhi default se kiya faisla ban jata hai. Aisa session jo teen din us approval par paused raha jo aap ne dekha hi nahin, zor se fail nahin hua. Woh khamoshi se fail hua, jo zyada bura hai.
Is liye aadat, aur kisi unattended cheez par bharosa karne se pehle yeh lazmi hai: gate ko jaan boojh kar aik baar ring karein. Chota task design karein jisay aik approval chahiye (misal ke tor par connected test folder mein harmless test file save karne ko kahein, aisa step jis ki tasdeeq ke liye woh rukta hai). Shuru karein. Computer se door chale jayein. Confirm karein ke notification waqai phone tak pohanchti hai, app installed hai, notifications on hain, focus mode ne dabai nahin, aur phone se approval clear karein. Aisa alarm jo aap ne kabhi suna nahin afwah hai. Project 3 is mein aap ki rehnumai karta hai.
Cowork remote session ke approvals Claude mobile app mein aap tak aate hain. Pehle walk-away task se pehle: app installed, usi account mein signed in, notifications enabled. Phir ooper ka deliberate-approval test aik baar shuru se aakhir tak chalayein.
OpenAI ne mobile gate seedha banaya. ChatGPT mobile app aap ko chalte kaam se connected rakhti hai. Desk se door aap sawalat ka jawab, outputs review aur aham actions approve kar sakte hain. OpenAI Work ko yun bayan karta hai ke woh goal ko steps mein torta hai jab ke aap saath chalte, sawalat ka jawab dete, simt badalte aur aham actions approve karte hain. Is liye gate design ke mutabiq phone tak aata hai. Cowork jaisi preflight: app installed, wohi account, notifications on. Phir kisi unwatched cheez par bharosa karne se pehle deliberate-approval test aik baar shuru se aakhir tak chalayein. (Kuch sab se tafseeli mobile-approval docs host Mac par Codex work ke liye likhi gayi hain. Work-mode ka exact behavior live pages par confirm karein. Magar gate-on-mobile khayal first-party hai.)
Worker aap ke door hone par kaam karta rehta hai, aur kabhi kabhi usay sawal poochna hota hai. Is surface par sawal phone par aata hai. Do rules. Pehla: phone se sawalon ka jawab dein, magar asal designing desk par karein. Doosra: arrangement par bharosa karne se pehle worker se jaan boojh kar aik sawal karwayein, aur check karein ke sawal waqai aata hai. Aisi doorbell jisay aap ne kabhi bajte nahin suna kharab ho sakti hai, aur aap ko kabhi pata nahin chalega.
Khud ko check karein
Aap ka scheduled Monday-morning brief do hafton se kuch nahin bana raha. Aap ko kabhi error nahin mila. Pehli aur doosri cheez kya check karenge?
Jawab dikhayein
Pehli: kya gate aap tak pohanchta bhi hai, kya session aise approval par paused hai jo phone tak nahin aaya? Gate ki silent non-delivery is concept ka failure mode hai, aur bilkul yeh symptom banati hai: na output, na error. Doosri: task chala bhi tha ya nahin, kyun ke metered usage aur plan limits (Concept 9) scheduled task ko pause, run se fail ya budget khatam hone par aap ki intervention ka mohtaj bana sakte hain, aur aap ko error milne par bharosa nahin karna chahiye. Ghor karein dono checks mein kya shared hai: is surface par khamoshi ambiguous hai jab tak aap usay koi matlab na dein. Yeh khayal engineering ki taqat ke saath Leaving the Laptop mein wapas aata hai.
Part 3: Baghair Nigraani Kaam
Aap surface jaante hain. Ab woh skill jo isay mufeed banati hai: aisa kaam hand over karna jisay aap nahin dekhenge. Yeh part delegation loop sikhata hai: woh chaar steps jinhein is kitab ka har baad ka course pehle se aana samajhta hai. Phir woh capability sikhata hai jis ke liye yeh surface bani: aisa kaam jo clock par shuru hota hai, jab aap ka koi device online na ho.
8. Delegation loop: brief, plan, approve, review
Yeh is kitab mein aap ka pehla delegation lesson hai, aur shayad zindagi ka bhi pehla. Foundations mein aap ne jo kuch kiya woh conversation tha: aap ne poocha, parha, dobara poocha, aur koi cheez aap ke dekhe baghair nahin hui. Delegation mukhtalif hai. Aap outcome bayan karte hain, agent multi-step kaam karta hai, aur aap nateeja dekhte hain. Nateeje ki quality taqreeban mukammal tor par chaar lamhon mein tai hoti hai, aur un mein se sirf aik prompt hai.
Step 1: Brief. Assignment us tarah bayan karein jaisay kisi smart, motivated naye colleague ko samjhayenge jo aap ka context nahin jaanta: outcome, constraints, audience aur wajah. Yeh Foundations Course 2 ka briefing discipline hai, stakes zyada hain, kyun ke ab model aap ke brief ke gaps ke khilaf sirf jawab nahin deta, un par action leta hai.
Farq aik pair mein:
- Query: "Is PDF ka khulasa karein."
- Assignment: "Yeh teen files parhein. Har deadline aur har open commitment flag karein. Mere manager ke liye aik-page brief banayein: seedha tone, deadlines pehle, koi introduction paragraph nahin."
Query chat mein theek kaam karti hai. Assignment is surface ke liye hai. Dekhein assignment mein query se zyada kya hai: kya banana, kis se, kis ke liye, kis shape mein, kin rules ke saath. In mein se har aik woh faisla hai jo warna agent aap ke liye khamoshi se, task ke darmiyan, aap ke baghair karta.
Step 2: Plan. Har non-trivial brief ko aik line par khatam karein:
Lay out your plan first, and pause for my approval before doing anything.
Yeh line is surface par har doosri jagah se zyada aham kyun hai. Desktop agent par aap aksar kaam chalta dekhte aur darmiyani step par rok sakte hain. Yahan aap door chale jayenge. Maqsad yehi hai. Is liye plan aam tor par intent aur finished work ke darmiyan aap ka wahid intercept hai. Isay chorein to ghalati pakarne ka pehla mauqa bohat dair se aata hai: ghalati task ki har file par chal chuki hoti hai.
Step 3: Approve ya redirect. Approve karne se pehle plan parhein. Chaar cheezein check karein, do minute lagte hain:
- Scope: kya yeh sirf aap ki batayi cheez chhoo raha hai, ya job khamoshi se barh gayi?
- Order: kya koi cheez verify hone se pehle action le rahi hai?
- Reach: kya yeh aisa connector ya send pesh kar raha hai jo aap ne nahin manga? (Agar haan, aur task kisi aur ke likhe content ko choota hai, to ruk kar Concept 6 dobara parhein.)
- Assumptions: format, audience ya rules ke baray mein yeh kya maan raha hai jo aap ne kabhi nahin kaha?
Agar kuch ghalat hai to zero se shuru na karein. Aik jumlay mein redirect karein: "Step 2 mein naye headings banane ke bajaye mojoodah template ke headings istemal karein." Agent plan dobara likh kar phir poochta hai. Plan stage par aik jumlay ki correction ki koi qeemat nahin. Pur-aitemad ghalat run ke baad cleanup aik dopahar leta hai.
Step 4: Review. Deliverable tab aata hai jab aap kahin aur hote hain, is liye review ko brief mein hi banayein. Har deliverable-producing assignment ke aakhir mein do lines shamil karein:
Before you start: ask me 1–2 clarifying questions.
If any sources contradict each other on a material point, flag the
contradiction in the deliverable. Do not silently pick one.
When done: list every file you created and where each one landed —
temporary working space, platform storage, or a system I control (connector save, download, local write, or repo commit).
Pehli line us assumption ko samne lati hai jo defect ban jata. Doosri is liye hai ke model ki training conflicts ko smooth karke pur-aitemad jawab banane ki taraf kheenchti hai, aur lawyer ya accountant ke liye do mutazad sources mein se khamoshi se aik chunna sab se aham failure hai. Teesri Concept 5 ka tier discipline hai jo automatic bana diya gaya. Phir nateeja parhte waqt: agar 90% durust hai to aakhri jumla haath se edit karein. Agar koi structural cheez ghalat hai to agent ko batayein aur revision maangein. Agent ke saath iterate karna usay aap ka taste sikhata hai. Haath se dobara likhna usay kuch nahin sikhata.
Apna pehla assignment ab chalayein (das minute, kuch connect nahin karna). Koi bhi real messy input lein, meeting notes, project plan, files mein paste ki teen related emails. Agent surface kholein, attach ya paste karein, aur poora four-step assignment bhejein: constraints aur audience wala outcome, plan-first line, teen review lines. Plan parhein. Jaan boojh kar aik cheez redirect karein, choti hi sahi, taa ke redirect kaam karta mehsoos ho. Approve karein. Chai jitni dair door chale jayein. Wapas aa kar deliverable ko brief ke mutabiq parhein. Yeh aik run baqi concept se zyada sikhata hai, aur bilkul wohi loop hai jo is ke baad har course mein baghair tabdeeli chalayenge.
Anti-pattern ka naam. Is surface par sab se aam failure isay superpowers wala chat box samajhna hai: unattended multi-step work par aik-line prompt phenk dena. Is ka hisaab sakht hai: brief jitna vague, agent akelay utne zyada faislay karta hai. Unwatched run jitna lamba, utne zyada faislay aap ke dekhne se pehle jama hote hain. Chat mein vague query aik reply zaya karti hai. Is surface par vague assignment aap ki aik dopahar cleanup mein zaya karta hai.
Har baar chaar steps. Jo chahiye poora batayein: result, rules, reader. Kaam shuru hone se pehle plan dekhne ko kahein. Plan parhein aur ghalat ho to aik jumlay se theek karein. Phir result ko apni request ke mutabiq parhein. Plan aham step hai. Plan ki fix aik minute leti hai. Mukammal ghalat kaam ki fix aik dopahar.
Mazeed gehrai: yeh loop choti shakal mein poora section hai
Aap ke aagay har course is loop ke aik step ko gehra karta hai. Cowork course desktop work ke steps 2 aur 3 ko gehra karta hai: per-step approval cards, execution view, poora trust model. Spec-Driven Development step 1 ko itna gehra karta hai ke brief aisi spec ban jata hai jis ke khilaf kaam check ho sake. Loop Engineering step 3 automate karta hai: checker agent kaam parhta hai taa ke aap har run na parhein. Trusting the Checker step 4 ko itna gehra karta hai ke review measurement ban jata hai. Baad ki koi learning is loop ko replace nahin karti. Har cheez is ke aik step ko tez karti hai.
Khud ko check karein
Aap multi-source task brief karte, plan-first line chorte aur meeting mein chale jate hain. Wapas aate hain to finished deliverable aik source ki ghalat reading par bana hai. Desktop agent par darmiyan mein pakar lete. Yahan woh safety net kyun kabhi dastiyab nahin tha, aur kaunsi aik line us ki jagah leti hai?
Jawab dikhayein
Net screen par aap ki aankhein tha, aur is surface par aankhein screen par honi hi nahin thin, kyun ke unwatched kaam maqsad hai. Aap ke liye mid-run intercept mojood nahin, is liye wahid intercept run se pehle hai: "Lay out your plan first, and pause for my approval." Is surface par woh line good practice nahin. Woh safety net hai.
9. Kisi device ke online hue baghair scheduled tasks
Ab tak har cheez aap ke start dabane se shuru hoti thi. Yeh concept usay bhi hata deta hai, aur yehi capability July 2026 ko aik course ke laiq banati hai.
Scheduled task woh assignment hai jisay aap cadence ke saath aik baar bayan karte hain, phir woh vendor ke servers par apni clock se chalta hai: laptop band, phone jaib mein, aap ka koi device online nahin. Har Monday 8 bajay haftay ka brief aap ke baithne se pehle tayyar hai. Wada yeh hai, aur is surface par haqeeqi hai. Yeh concept woh discipline shamil karta hai jo isay khamoshi se mess banne se rokta hai.
Har scheduled task chaar jawabat hai. Banane se pehle isi order mein likhein:
- Isay kya karna chahiye? Standing brief. Concept 8 ki har baat, aik izafay ke saath: clarify karne ke liye koi nahin hoga, is liye brief ko aap ki ghair-mojoodgi mein bhi chalna chahiye, empty case samet. "Agar is haftay koi naya item nahin, to aik-line note banayein jo yehi kahe" quiet week aur aisi silent failure ka farq hai jinhein aap alag nahin pehchan sakte.
- Yeh kya chhoo sakta hai? Scope ki files aur platform storage, naam se. Vague scope wala scheduled task har haftay zyada wasee hota hai, aur koi isay drift karta nahin dekh raha.
- Yeh kahan pohanch sakta hai? Connectors ki list. Connectors permissions hain, suggestions nahin: mail connector ke baghair task mail nahin bhej sakta chahe brief kuch bhi kahe, aur yeh feature hai. Concept 6 ki tarah task ko job ke liye sab se choti rasai dein, aur kuch nahin, kyun ke yeh grant baghair nigraani schedule par ghair-muayyan muddat tak istemal hogi.
- Yeh kab shuru hota hai? Cadence. Aur har banay scheduler ki aik imaandaar haqeeqat: "Monday 8 bajay" ka matlab Monday 8 ke qareeb hai. Schedulers stagger aur kabhi slip karte hain. Agar output se hard deadline juri hai to schedule akela guarantee nahin, jis ko Leaving the Laptop durust tarah samjhata hai.
Pehle schedule se pehle teen rules, har aik ki qeemat kisi aur ne mushkil se di:
- Us cheez ko kabhi schedule na karein jis par door jane ka bharosa pehle se nahin. Schedule seerhi ki aakhri pairi hai, pehli nahin. Task ko kam az kam chand baar haath se aur dekhte hue chalayein. Kuch baar aur baghair dekhe chalayein. Sirf door jane mein kamyab cheez clock ki mustahiq hai. Aik baar chalaye task ko schedule karna automation nahin. Yeh weekly surprise ki subscription hai.
- Metered hisaab likh kar karein. Kisi aur ke servers par hands-off work ka hamesha budget hota hai: roz ke runs ya har plan ke hours, tier ke mutabiq staged aur har company ke liye mukhtalif. Jis haftay scheduled tasks takrayein us se pehle unhein plan limit ke khilaf ginein. Usage ya task limits poori hon to task pause, run se fail ya aap ki intervention ka mohtaj ho sakta hai. Error message na hone par bharosa na karein. Scheduled page aur notifications check karein. Yehi failure Concept 7 ne bataya: khamoshi ka koi matlab nahin jab tak aap usay matlab na dein.
- Run ka complete hona task ki kamyabi nahin. Scheduler ka kaam "yeh chala" par khatam hota hai. Kaam durust hai ya nahin scheduler ka masla nahin. Is liye success signal task ke output mein banayein: brief aik wazeh line, processed items, bani files aur un ke tiers, ya explicit empty-case note par khatam ho, aur aap woh line parhein. Green ka matlab "chala" hai. Sirf line ka matlab "kaam hua" hai.
Cowork ke scheduled tasks cadence ke saath saada zaban mein describe hote hain, aur web par remotely chalte hain: aap ke kisi device ke online hue baghair fire hote aur results session mein intezar karte hain. Access aur run limits plan ke mutabiq staged aur metered hain, hisaab apne plan ki live limits se karein, yahan printed cheez se nahin. Aik caveat jaldi jaan lein, aur woh limit ka ulat hai: Cowork scheduled tasks remotely chalte hain, computer so raha ho ya desktop app band ho tab bhi. Woh connectors, plugins aur Claude account mein saved files ke saath kaam karte hain, magar computer ke folder se bandhe nahin ho sakte. Is liye recurring workflow local files ya desktop apps par depend kare to mukhtalif, desktop-dependent setup chahiye. Scheduled work ko connectors aur account storage par rakhein to wada qaim rehta hai.
ChatGPT Work ke Scheduled Tasks OpenAI ke apne alfaaz mein seekhay chaaron heartbeats cover karte hain: aik action aik baar, schedule ya event par repeat, ya waqt ke saath tabdeeli monitor. (Is web surface par "nayi file nazar aaye" jaisay event-style triggers instant push ke bajaye periodic monitoring se handle hote hain, is liye yahan "trigger" aur "monitor" ko aik family samjhein.) Launch examples mein dono qismein hain. Kuch report karte hain: har subah dashboards check aur tabdeeliyon ka khulasa. Doosray action lete hain: naya feedback aane par presentation update, ya recurring feedback themes ko prioritized product ideas mein badalna. Product report se zyada karta hai. Yeh beginner course phir bhi jaan boojh kar reporting line par rukta hai. Action schedules ka blast radius bara hai (zyada cheezein ghalat aur door tak ja sakti hain), aur baad ke courses mein checker design, state, idempotency aur recovery seekhne tak intezar karna chahiye. Yeh course ka chuna rule hai, product ki limit nahin. Chaar answers (what, touch, reach, when) aur teen rules yahan baghair translation lagu hote hain. Tasks account ke khilaf cloud mein chalte hain. Reach Plugin Directory se grant hoti hai. Usage plan ke agent allowance se kam hoti hai, is liye metered hisaab yahan bhi haqeeqi hai. Mojoodah allowance live pages par confirm karein.
Aik saaf caveat: aise Project ke andar banaya Scheduled Task jis mein uploaded files hon, un Project files tak nahin pohanch sakta. Is liye supported connector ya aisa doosra source istemal karein jahan task pohanch sakay, aur scheduled workflow ko Project mein uploaded files ke gird design na karein. Yeh Concept 3 amal mein hai: wohi six-part shape, mukhtalif company ke buttons.
Aur boundary, imaandari se. Sirf is concept se aap reporting schedule bana sakte hain: sources parhein, synthesize karein, brief banayein, durust tier mein utarein. Yeh waqai qeemti hai, is surface par aksar logon ki pehle saal ki value yehi hoti hai, aur jaan boojh kar is course ki had hai. Abhi jo mehfooz nahin bana sakte woh aisa schedule hai jo action leta hai: cadence par send, file, update aur decide karta hai. Baghair nigraani action ko woh parts chahiye jo abhi aap ke paas nahin, kaam grade karne wala checker, machine se verify honay wali stopping condition, runs ke darmiyan yaad rakhne wali state file. Loop Engineering yeh sab design karna sikhata hai, aur Leaving the Laptop proven result kahan rehna chahiye sikhata hai. Is concept ne darwaza khola. Woh courses us mein se guzarte hain.
Scheduled task standing instruction hai: worker har Monday baghair poochay job karta hai, laptop off ho tab bhi. Chaar cheezein likhein: job, woh kya chhoo sakta, kahan pohanch sakta, aur kab chalta hai. Teen rules: sirf woh jobs schedule karein jo worker aap ke door hone par pehle hi achi kar chuka. Limit tak pohanchne se pehle plan ka budget ginein, aur khamoshi par bharosa karne ke bajaye Scheduled page check karein. Aur job ko apna result report karwayein, kyun ke "yeh chala" aur "yeh kaam kar gaya" mukhtalif facts hain.
Khud ko check karein
Ayesha do cheezein schedule karna chahti hai: clients ke unpaid invoices ka Monday-morning summary (Drive se read, brief account mein save), aur automatic Friday task jo har late client ko payment reminder email karta hai. Kaunsa is course mein hai, kisay intezar karna chahiye, aur dono mein aik-lafz ka farq kya hai?
Jawab dikhayein
Monday summary yahan hai: reporting schedule, read, synthesize, tier mein land, aur ghalat hafta sirf rewrite ki qeemat rakhta hai. Friday reminders ko intezar karna chahiye: woh task bahar ki duniya par baghair nigraani cadence ke saath action leta hai, ghalat run real clients ko email karta hai, aur is course mein koi cheez bahar jane se pehle kaam check nahin karti. Aik-lafz ka farq: acting (reporting ke muqablay mein). Reporting schedules is course ki had hain. Acting schedules ko Loop Engineering ka checker, stopping condition aur state chahiye, aur tab bhi reminder email bilkul woh kaam hai jis ke liye human gate mojood hai.
Part 4: Chunao aur Open Path
Ab aap is surface par agent ko direct, files ko durust tier mein land aur reporting task ko clock par rakh sakte hain. Teen sawal reh gaye, aur woh imaandaar sawal hain. Yeh surface kab ghalat chunao hai? Vendor cloud ke baghair open-source path kaisa hai? Aur products kitne bhi achay ho jayein, yeh surface aap ke liye kya kabhi nahin kar sakti? Yeh part teeno ka jawab deta aur phir aap ko baqi section ke hawalay karta hai.
10. Web, desktop ya terminal: kaam jis cheez ko choota hai us se chunein
Jo surface abhi seekhi woh kai options mein se aik hai, aur mature aadat aik se wafadari ka ahad nahin, har task ke liye chunao hai. Har case ko sort karne wala rule chota hai: kaam jis cheez ko choota hai us se chunein.
| Kaam ko chahiye... | Surface | Yeh kitab kahan sikhati hai |
|---|---|---|
| Connector-and-document work, devices mein continuity, kuch install nahin | Web | Yeh course |
| Aisa kaam jo machine band hone par jari rahe | Web (repo work ke liye cloud Routine, baad ka topic) | Yeh course; Loop Engineering |
| Aap ki local files, desktop applications, woh browser jahan aap signed in hain | Desktop | Cowork & OpenWork, knowledge workers ke liye agla |
| Code, repositories, command line | Coding agents | Agentic Coding, engineers ke liye agla |
| Custody requirements wala regulated data (PHI, privileged matter, regulated financial data) | Abhi ooper mein se koi nahin. Pehle compliance ka jawab likh kar lein | Cowork course mein regulated-workloads ki poori rehnumai hai |
Table do baar parhein, kyun ke yeh do kaam karti hai. Task router ke tor par aaj ka sawal hal karti hai: yeh report, yeh surface. Course router ke tor par section ka sawal hal karti hai: agla co-worker course kaunsa. Jin knowledge workers ka din documents aur connectors mein rehta hai woh Cowork course par jate hain, jo web ki rasai se bahar har cheez shamil karta hai. Engineers Agentic Coding par jate hain. Kisi ko yeh surface chorne ki zaroorat nahin: sehatmand setup hamesha tasks ke mutabiq surfaces milati hai.
Aik row apne paragraph ki mustahiq hai kyun ke is mein sab se zyada risk hai: regulated data. Concept 5 yaad karein: tiers 1 aur 2 definition ke mutabiq vendor ki custody hain. Remote session ki working files aur platform storage vendor ki machines par rehte hain. Isi se surface kaam karti hai. Aur bilkul isi par compliance regime tawajju deta hai. Is liye PHI, privileged client matter ya regulated financial data ke liye sawal "kya yeh surface ehtiyat karti hai?" nahin. Asal sawal hai "kya yeh data is custody mein, is plan aur is paperwork ke tehat reh sakta hai?" Aur sirf compliance team likh kar jawab de sakti hai. Cowork course poori tafseel rakhta hai. Jab tak likha jawab na mile, regulated data is surface ko nahin choota.
Multi-surface mix ko mess se bachane wala aik rule aap pehle hi jaante hain: jaanein har deliverable kis tier mein land hua. Tier 3, aap ke Drive, repo ya firm records, mein deliverable ko farq nahin parta kis surface ne banaya. Tier 3 ka poora maqsad yehi hai.
Mukhtalif jobs, mukhtalif workplaces. Documents aur connected services, ya aisa kaam jo machine band hone par bhi chale: web. Apni files aur desktop programs: aglay course ki desktop apps. Code: coding agents. Sensitive regulated data: rukein, pehle compliance se likha haan lein. Aik waqt mein kai workplaces istemal karein, bas yaqeen rakhein ke har finished document aap ki filing cabinet mein khatam ho, jis workplace ne bhi banaya.
Khud ko check karein
Teen tasks desk par aate hain: (a) Slack aur Notion se weekly summary jo safar mein bhi chahiye. (b) laptop ke folder mein 200 files clean karna. (c) privileged client documents ko synthesize karta memo. Har aik ko route karein aur deciding fact batayein.
Jawab dikhayein
(a) Web: yeh sirf connectors ko choota aur machine band hone ke baad bhi chalna chahiye, dono aik taraf ishara karte hain. (b) Desktop: kaam local filesystem ko choota hai, jahan browser-only remote session apne bal par seedha nahin pohanch sakta. Yeh Cowork course hai. (c) Abhi dono nahin: privileged matter custody ka sawal hai, aur tiers 1 aur 2 definition se vendor custody hain, kisi surface ke data dekhne se pehle compliance likh kar jawab deta hai. Teen tasks, teen answers, aik rule: kaam kis cheez ko choota hai woh faisla karta hai.
11. Open path: vendor cloud ke baghair
Is course ki structure ke baray mein aik iqrar, phir woh imaandari jo isay darkar hai.
Is kitab ka har co-worker course vendor tool ko open-source twin ke saath rakhta hai: Claude Code ke saath OpenCode, Cowork ke saath OpenWork. Is course ne woh pattern tora. Is ke tabs do closed products ko saath rakhte hain: web par Cowork aur ChatGPT Work. Yeh jaan boojh kar tha. Do rivals ne chand din ke farq se aik shape release ki. Aik ne remote session aur doosray ne chat app ke andar mode ke tor par. Yeh is kitab ki thesis ka saboot hai, aur aap ko side by side dekhna tha. Magar khamoshi se tora pattern bias ban jata hai. Is liye woh sawal jis ka jawab missing open-source tab deta: kya vendor cloud ke baghair yeh surface mil sakti hai?
Haan, mehnat ke saath. Aaj do open paths hain. In ke kuch alfaaz (self-hosted, cron, GitHub Actions) engineers ke hain; abhi aap ko unhein jaanne ki zaroorat nahin, sirf itna ke yeh paths mojood hain.
OpenWork, remote aur shared workspaces. OpenWork open-source desktop co-worker hai jis se aglay course mein theek se milenge. Yeh laptop ke ilawa kahin chalne walay worker se connect ho sakta hai. Woh self-hosted OpenWork worker ho sakta hai jahan URL aur access token se pohanchein. Ya organization ke shared cloud workers jin se aik step mein connect hon. Yeh is course ka bunyadi khayal hai: kaam karne wali machine woh nahin jo aap dekh rahe hain. Magar yahan yeh aap ya organization ke control walay infrastructure par chalta hai. Window aap ki hai. Building bhi.
OpenCode, aap ke apne scheduler ke saath. Repo-attached work ke liye Concept 9 ka open jawab aap ki clock se chalaya coding agent hai: aap ki machine par cron, ya scheduled GitHub Actions job. Na vendor cloud, na plan tiers, na staged rollout. Poora pattern Loop Engineering sikhata hai. Aaj sirf jaanein ke yeh mojood hai aur ajeeb nahin.
Ab trade, aik baar aur seedha. Is kitab mein har open-versus-managed chunao isi par aata hai. Companies aap ko spine bechti hain. Open path aap se spine banwata hai. Vendor surface par har cheez pehlay din working milti hai: sessions, files, sync, phone gate, scheduler. Magar vendor ki custody, format, qeemat aur rules mein. Open path par har cheez aap set up, run aur fix karte hain. Badlay mein do cheezein milti hain jo koi vendor nahin bechta. Pehli, custody: aap ka data aap ke control ki machines par. Doosri, choice: kaunsa model prompts dekhe, aur switch karne ki azadi. Trade ka koi pehlu nek nahin. Concept 10 ki regulated-data row mein custody seedha jeetti hai. Friday tak briefs ship karta solo consultant woh case hai jahan working spine seedha jeetti hai. Zyada tar readers har task ke liye alag, dono jawab saath rakhenge. Ab tak yeh jana pehchana lagna chahiye.
Vendor surface furnished office hai: desk, drawers, doorbell, pehlay din sab working, un ki building mein. Open path khali kamra hai jisay aap khud furnish karte hain: zyada kaam, magar building, rules aur kis ko kya nazar aaye ka chunao aap ka. Koi sirf behtar nahin. Sensitive work ko aksar aap ki building chahiye. Tez work ko aksar furnished office. Aap aik hi waqt mein aik rent aur aik own kar sakte hain.
Khud ko check karein
Aik NGO weekly donor-report automation set up karwana chahti hai. Data policy donor records ko third-party consumer platforms par mana karti hai, magar team mein engineer nahin. Policy kaunsa path lazmi karti hai, team kya khoti hai, aur commit karne se pehle aap ko kaunsa imaandaar jumla kehna chahiye?
Jawab dikhayein
Policy open path lazmi karti hai: donor records consumer vendor platform ke tiers 1 aur 2 mein nahin utar sakte, aur woh tiers vendor surface ki bunyaad hain. Woh free spine, sync, phone gate aur managed scheduler khote hain, jinhein ab kisi ko banana aur chalta rakhna hai. Imaandaar jumla: "Yeh path subscription fee ke badlay operator leta hai. Yahan kisi ko setup, updates aur us subah ki zimmedari leni hogi jab yeh tootay. Aur staff mein engineer na ho to software nahin, woh qeemat asal faisla hai." Custody muft nahin. Is ki qeemat operations mein di jati hai.
12. Yeh surface kya nahin kar sakti, aur raasta kahan jata hai
Course ko imaandaar boundary par khatam karein, jaisay is kitab ka har course hota hai.
Surface kaam behtar nahin karti. Remote session kamzor brief ko behtar nahin, unattended banata hai. Har quality lever screen ki aap wali taraf hai: brief (Concept 8), plan review (wahid intercept), tier decision (Concept 5), deliberate gate test (Concept 7), walked-before-scheduled rule (Concept 9). July 2026 ne agent ko browser mein dala. Aap ki koi zimmedari kahin nahin gayi.
Surface chand haftay purani, metered, staged aur vendor-shaped hai. Is page ka har mechanical fact, plan tiers, run limits, button names, hatta ke product names, kitab ka sab se jaldi aging material hai. Lasting layer saath le jayein: stop-typing test, window-not-runtime architecture, teen tiers aur aik-line discipline, six-part reading lens, four-step loop, schedule ke chaar answers. Yeh har rename ke baad qaim hain. Aur aik discipline aap ko sari churn se bachata hai, is liye aakhri baar sunein: koi aham cheez sirf platform par nahin rehti. Jis cheez ko khona bura lagay us ke liye hamesha Tier 3.
Aur yeh surface shuruat hai, manzil nahin. Ab aap is section mein chalne ke qabil hain, aur har course aaj ki learning mein yeh shamil karta hai:
- Cowork & OpenWork, knowledge workers ke liye agla, delegation loop ko real files aur desktop tak barhata hai, poore trust model ke saath jis ki web ko zaroorat nahin thi: folders as a lever, autonomy ladder, regulated-workloads discipline.
- Agentic Coding, engineers ke liye agla, wohi loop terminal mein code par.
- Spec-Driven Development aap ke step 1 ko tez karta hai jab tak brief aisi spec na ban jaye jis ke khilaf kaam check ho, aur yeh surface par bhi chalta hai.
- Loop Engineering Concept 9 ki roki cheez banata hai: action lene walay schedules, checker, stopping condition aur state spine ke saath jo action ko mehfooz banate hain.
- Harness Engineering aur Trusting the Checker loop ko pehle mehfooz, phir measurable banate hain.
- Leaving the Laptop runtime decision ke saath section band karta hai, proven work kahan rehna chahiye, aur aaj seekhi surface is ke gharon mein se aik ka front door hai.
Aakhri khayal, aur teen jumlon mein poora course. Chat box mein aap design karte hain. Agent surface par kaam chalta hai. July 2026 mein dono aik address par aa gaye, aur kaam ki taqseem kamzor nahin zyada saaf hui. Shape seekhein. Address phir badlega.
Yeh surface badalti hai ke kaam kahan aur kab chalta hai. Yeh kaam ki quality nahin badalti: woh ab bhi brief, plan review aur checks se aati hai. Is page ke products kai baar names aur limits badlenge. Is page ke khayalat nahin. Finished work apni storage mein rakhein, aur khayalat aglay course mein le jayein.
Khud ko check karein
Reader course khatam karke kehta hai: "Ab agents web par chalte hain, desktop aur terminal courses legacy hain." Concept 10 ke rule aur is concept ki boundary se do-jumlon ki correction likhein.
Jawab dikhayein
Surfaces is se taqseem hoti hain ke kaam kis cheez ko choota hai, is se nahin ke aakhir mein kaunsi release hui. Local files, desktop apps aur repos haqeeqi kaam hain jahan browser-only remote session apne bal par seedha nahin pohanch sakta. Is liye desktop aur terminal courses woh ilaqa sikhate hain jo yeh surface cover nahin karti. Aur koi surface, nayi ya purani, kaam khud behtar nahin karti: brief, plan review aur checks skill hain, aur baqi section mein woh skill gehri hoti hai.
Yeh kitab yeh surface kahan istemal karti hai (dogfooding)
Mid-July 2026 tak imaandaar jawab: is kitab ke apne unattended loops is surface par nahin chalte, aur wajah Concept 10 ka kaam hai. Kitab repository hai, loops repo-attached hain, aur review gate CI mein merge gate ke paas rehta hai, bilkul jaisay Leaving the Laptop ka dogfooding section samjhata hai. Kaam jis cheez ko choota hai, repo, us ne doosri surface par route kiya.
Is kitab ki production mein yeh surface do jagah apni qeemat dikhati hai. Pehli, har loop se pehle chat-side design work: briefs draft, reviewer prompts test, aik beat haath se run. Doosri, kitab ke gird connector-and-document tasks. Doosri qisam reporting work hai jisay Concept 9 ne course ki had kaha. Yeh taqseem compromise nahin. Course ki sikhai division of labor hai: chat box mein design, us surface par run jisay kaam choota hai. Aap abhi isi arrangement ka output parh rahe hain.
🚀 Projects
Chhay projects, aasan se mushkil. Real magar low-stakes kaam, notes, plans aur drafts istemal karein, regulated data kabhi nahin (Concept 10 ka rule practice par bhi lagu hai).
📱 Project 1 · 15 min · Closed-lid test. Difficulty: easy · Uses: Concept 2. Karein. Chota multi-step session shuru karein (teen pasted documents ko aik brief mein summarize). Run ke darmiyan tab band karein. Phone se session dobara kholein. Tab mukammal jab aap ne wohi session doosray device par jari dekha aur aik jumlay mein bata sakein ke runtime asal mein kahan hai.
🗂️ Project 2 · 30 min · Three-tier audit. Difficulty: easy · Uses: Concepts 5 and 8. Karein. Aik real task chalayein jo deliverable banaye. Concept 8 ka poora brief istemal karein, aakhri line samet jo har file aur tier list karti hai. Tab mukammal jab task ki har file ka tier bata sakein, aur deliverable khud tier 3, aap ke Drive, inbox ya records, mein ho, sirf platform par nahin.
🔔 Project 3 · 20 min · Gate ko jaan boojh kar ring karein. Difficulty: easy–medium · Uses: Concept 7. Karein. Aisa task design karein jisay aik approval chahiye, phir computer se mukammal door chale jayein. Approval behavior product, connector aur workspace ke mutabiq badalta hai, is liye safe, low-stakes trigger istemal karein: agent se connected test folder mein harmless test file save karwayein, ya aisa draft action tayyar karwayein jisay aap ki tasdeeq chahiye. Is drill mein sending, deleting, purchasing ya account changes na karein. Agar approval fire na ho to woh bhi finding hai: approval mode aur connector settings check karein, aur note karein gate ring karne mein kya laga. Tab mukammal jab approval phone tak aaye, aap phone se clear karein, aur wapas aane par finished work intezar kar raha ho. Aisa alarm jo kabhi suna nahin afwah hai. Yeh project aik baar sunwata hai.
⏰ Project 4 · 30 min, plus aik hafta · Pehla reporting schedule. Difficulty: medium · Uses: Concept 9. Karein. Aik connector aur aik file source se Monday-morning brief banayein. Chaar answers pehle likhein (what, touch, reach, when), brief mein empty-case line aur success signal shamil karein, aur plan ka metered hisaab likh kar karein. Aik vendor note: ChatGPT par uploaded files walay Project ke andar banaya Scheduled Task un Project files tak nahin pohanch sakta. Supported connector ya doosra reachable source istemal karein, aur scheduled workflow Project mein uploaded file ke gird na banayein. Tab mukammal jab aap ka koi device online hue baghair do baar fire ho, aur har result mein sirf completed run nahin, success signal ho.
🔁 Project 5 · 45 min · Aik task, do vendors. Difficulty: medium · Optional: dono products ki access darkar · Uses: Concept 3. Karein. Wohi assignment web par Cowork aur ChatGPT ke Work mode mein chalayein. Dono taraf paid plan chahiye, aur staged rollouts ki wajah se shayad dono tak na pohanchein. Aisa ho to doosray product walay shakhs ke saath results compare karein, ya apne product par chala kar Concept 3 ki table se doosray ka andaza lagayein. Concept 3 ki comparison table paas khuli rakhein aur test karein: kya har row ab bhi durust hai, ya mid-July 2026 ke baad rollout ya rename ne kuch badla? Tab mukammal jab tier 3 mein saved aik-page note ho. Woh bataye: comparison ki kaunsi rows confirm hui. Kaunsi drift hui (mechanical layer real time mein aging). Aur kya chhay shared parts dono mein nazar aaye. Aakhri finding maqsad hai, differences nahin.
🧳 Project 6 · 45 min · Exit drill. Difficulty: capstone · Uses: Concepts 5 and 11. Karein. Tasawwur karein platform is quarter gaib ho jata hai. Sirf tier 3, Drive, email, records, se working state dobara banayein: templates, standing briefs, deliverables. Tab mukammal jab jo cheez dobara nahin ban sakti us ki list likhi ho. Woh list tier-2 exposure aur aglay haftay ka saving discipline hai. (Leaving the Laptop par yeh drill engineering ki taqat ke saath dobara milegi.)
Appendices: baad ke courses, web par
General Agents section ka baqi hissa apna craft Claude Code aur OpenCode se sikhata hai: local tools, terminal se chalte, aap ki machine ki files par kaam karte. Is course ne web surface sikhai: web par Cowork aur ChatGPT Work, dono remote, dono browser tab mein.
Yeh appendices pul hain. Har baad ke course ke liye aik sawal poochte hain: bina kuch install kiye web surface se us course ka kitna kaam ho sakta hai? Jawab kabhi "sab" aur kabhi "kuch nahin" nahin. Kuch khayalat mukhtalif names ke tehat saaf transfer hote hain. Kuch kamzor shakal mein. Aur kuch bilkul paar nahin ja sakte, kyun ke local machine ya real code execution chahiye jo web surface nahin deti. Aakhri appendix course se peeche hat kar poori industry mein harness war map karta hai.
Woh aakhri group web surface ka gap nahin. Yehi woh lakeer hai jahan section browser chorne ko kehta hai. Appendices do tareeqon se parhein: aaj jo kar sakte hain us ka map, aur is ka preview ke aglay courses kyun mojood hain.
Sab mein aik rule qaim hai. Web par aap finished harness rent karte hain. Baad ke courses mein banate hain. Topic tab "transfer" hota hai jab vendor ka harness woh kaam pehle hi aap ke liye karta hai, aur tab "transfer nahin" hota jab aap ko khud wire karna ho. Lakeer samne rakhein to har appendix aik hi tarah parhta hai.
Aik aur framing note taa ke boundaries imaandaar rahein. Jab appendix kehta hai web "local files tak nahin pohanchta" ya "local repo nahin", matlab browser-only session hai: aik tab, kuch installed nahin. Course ne yehi surface sikhai. Is se zara aagay do cheezein hain. Desktop bridge (Concept 2) desktop app khuli ho to remote session ko mahdood local access de sakti hai. Aur local agent (desktop aur coding courses) shuru se machine par chalta hai. "Does not transfer" rows ko browser-only parhein: yeh tab ka aakhir dikhati hain, bilkul jahan aglay courses shuru hote hain.
Appendix A · Agentic Coding, web par
Agentic Coding course Claude Code aur OpenCode ko pandrah concepts se sikhata hai. Yeh kitab ka context-engineering course bhi hai. Zyada tar khayalat model kya dekhta hai usay manage karne aur kaam checkable banane ke hain. Hairat angez hissa web par transfer hota hai, kyun ke chhay parts wohi hain. Jo nahin hota woh terminal, local file system aur khud chalaye code ko assume karta hai.
Saaf transfer hota hai:
| Woh course sikhata hai | Web par yeh hai |
|---|---|
| Plan mode (pehle parhein, approval ke baad action) | Concept 8 ka delegation loop. Run se pehle plan maangein. Wohi khayal, koi keybind nahin. |
| Rules file (CLAUDE.md / AGENTS.md), permanent project instructions | Project ki saved instructions, aur Concept 6 ka persistent-workspace khayal. Standing rules jo agent har session parhta hai. |
| Context management (window lean, fresh shuru ya continue) | Fresh session ya continue karne ka chunao. Concept 4 ki state spine wohi stack hai, aap ke liye managed. |
| Model ko task se milayein | Web par patla: surface ka model milta hai, menu nahin. Aadad "jaanein yeh surface kaunsa model chalati hai" ke tor par rehti hai. |
| MCP, system of record connect karna | Bilkul connectors. Concept 6. Neeche wohi standard, dono taraf mukhtalif catalog. |
| Skills (packaged, reusable know-how) | Eligible plans par dono families chat se Skill bana, editor mein build, ya instructions, resources aur code wala package upload karne deti hain. Format transfer hota hai. |
| Verification loop (attempt, check, fix, repeat) | Scheduled report jo condition check karke tabdeeli batati hai. Concept 9 ki had. Check aap ke liye chalta hai. Abhi auto-fix nahin. |
| Prompt injection, context darwaza hai | Concept 6 ka untrusted-content rule, lafz ba lafz. Risk aur aadat har surface par aik hain. |
Kamzor shakal mein transfer hota hai:
| Woh course sikhata hai | Web par mahdood version |
|---|---|
| Slash-commands aur Skill ke gird runtime | Skills khud transfer hoti hain (ooper dekhein). Kamzor coding-agent setup rehta hai: authored slash-commands, aur Skill ke runtime, lifecycle hooks, model routing, filesystem aur enforcement layer par aap ka control. Skill milti hai, chalne wali machine nahin. |
| Subagents (apni context window wala alag worker) | Web surface bara task khud sub-steps mein phaila sakti hai, magar aap chosen model aur restricted toolset wala subagent author nahin karte. Vendor ki orchestration milti hai, apni nahin. |
| Resume, undo, rewind | Sessions resume hote hain. Magar pichlay step ke liye /rewind aur agent edits ka git-backed /undo nahin, kyun ke kaam repo mein nahin hua. Undo hai "purani file tier 3 mein rakhein." |
Transfer nahin hota, aur yehi nuqta hai:
| Woh course sikhata hai | Web kyun nahin kar sakta |
|---|---|
| Hooks aur plugins (code mein enforce rule, har baar) | Hook aap ki machine par action block ya check karne wala code hai. Web surface aap ka code nahin chalati ya lifecycle events expose nahin karti. Yeh poore course ki sab se saaf "browser chorein" lakeer hai. Jab aisi guarantee chahiye jis se model baat karke na nikal sakay, aap ko apna banaya harness chahiye. Woh Harness Engineering hai. |
| Local repo par kaam, terminal commands | Browser-only session company servers par, disk par nahin chalta. Apne bal par terminal nahin kholta ya local project edit nahin karta. Desktop bridge mahdood access de sakti hai, magar terminal-and-repo workflow local tools ka ghar hai. Concept 2 dobara: runtime remote hai. |
| Models route karke cost control (frontier model par plan, cheap par execute) | Plan-with-strong, execute-with-cheap taqseem ko do aap ke chosen models aur mid-task swap wala tool chahiye. Web surface model aap ke liye chunti hai. |
Imaandaar khulasa. Agar kaam documents, connectors aur scheduled reports hai to web surface mukhtalif names ke tehat Agentic Coding course ka zyada hissa deti hai. Jis lamhay kaam apne code, files ya code mein enforced rule ko choota hai, web ka kinara mil gaya. Kinara flaw nahin. Aglay courses ka darwaza hai, aur ab aap jaante hain woh kahan hai.
Appendix B · Cowork aur OpenWork, web par
Yeh baqiyon se mukhtalif hai. Cowork & OpenWork course is course ka qareebi rishtay dar hai. Wohi qisam ka general-agent co-worker, unhi knowledge workers ke liye, wohi delegation loop. Farq runtime hai. Woh course aap ke desktop par do products sikhata hai: Claude Desktop ke andar Cowork, aur open-source twin OpenWork. Yeh course Cowork ko web par sikhata hai. Wohi family, alag ghar. Is liye map "kitna transfer" nahin. Map hai "desktop kya deta hai jo web nahin deta, aur kyun."
Saaf transfer, kyun ke product family wohi hai:
| Woh course sikhata hai | Web par yeh hai |
|---|---|
| Delegate karein, query nahin (co-worker ko assignment) | Concepts 1 aur 8, lafz ba lafz. Sab se aham aadat, har surface par aik. |
| Task loop (brief, plan, approve, execute, review) | Concept 8 ka delegation loop. Wohi chhay steps, wohi plan par intercept. |
| Plan leverage hai | Concept 8 ka core, aur delegation-loop figure plan step highlight karti hai. |
| Trust lever ke tor par connectors (sab se chota scope, read send nahin) | Bilkul Concept 6. Wohi connectors, scope discipline aur "read scope send scope nahin" rule. |
| Approval modes (action se pehle poochein, baghair poochay action) | Concept 7 ka phone gate. Web wohi approval inline card ke bajaye phone par route karta hai. |
| Skills (description se agent ka chuna saved playbook) | Web par Cowork wohi Skills system rakhta hai: catalog, /skill-creator, aur SKILL.md format product ke hain, desktop ke nahin. Yahan web ooper ke appendices ke ishare se mazboot hai. |
| Autonomy ladder (har task type par aik rung) | Concepts 8 aur 9: jis se door na jayein usay schedule na karein. Ladder trust ki hai, surface ki nahin. |
| Prompt injection, untrusted content | Concept 6 ka rule. Chhupi instruction wali vendor-PDF story web par aik jaisi hai. |
| Scheduled tasks (Cowork ki cadence) | Concept 9. Cowork ka /schedule aur Scheduled view web se wohi kaam karte hain. |
| Layered instruction stack (global, folder, session) | Concepts 4 aur 6: global settings, Project instructions, session prompt. "Folder" layer Project layer ban jati hai. |
Transfer hota hai, magar web shape badalti hai:
| Woh course sikhata hai | Web par farq |
|---|---|
| Teen trust levers: folders, connectors, approvals | Do aik hain. Pehla badalta hai. Desktop par local folder ki access dete hain. Browser-only session par local folder nahin. Working area Concept 5 ka task filesystem aur platform storage hai. Scope discipline rehti hai, magar "disk par dedicated working folder" "kaam apne Project mein rakhein" banta hai. (Desktop bridge local folders tak ja sakti hai, magar woh app ka kaam hai, tab ka nahin.) |
| Sub-agents (batch par parallel workers) | Web par Cowork batch ko khud parallel workers mein phaila sakta hai. Aap har worker ki desktop wali poori visibility aur local files par dependent control khote hain. Pattern rehta hai, local-file version nahin. |
Persistent workspaces (folder + CLAUDE.md) | Concept 6 ka Project + saved instructions. Wohi khayal, disk par folder nahin. Context file Project ka instruction field ban jati hai. |
| Model routing (sochne ke liye strong, plumbing ke liye cheap) | Appendix A jaisa patla. Har task surface ka model picker, OpenWork ka poora provider choice nahin. |
Transfer nahin hota, aur desktop ka poora maqsad yehi hai:
| Woh course sikhata hai | Web kyun nahin kar sakta |
|---|---|
| Execution surface ke tor par local files | Desktop course ka dil. Agent aap ki machine ke folder mein files read/write karta hai: matter folder, close-cycle folder, campaign folder. Remote web session company ke servers par aur apne bal par local drive se door. Yeh Concept 2 aur alag desktop course ki sab se bari wajah hai. |
| OpenWork khud (open-source, bring-your-own-model twin) | OpenWork install honay wali desktop app hai, aap ke chosen provider ke saath, files machine par. Is trade ka web version nahin. Open path baad ka topic hai, Concept 11 naam deta hai, magar browser surface nahin. |
| "Local-first" privacy (files machine se na niklein) | Web par agent ka parha file content definition se company servers par hai. Desktop ki core privacy story remote surface par nahin ho sakti. Regulated data ke liye faisla kun, Concept 10 lakeer kheench chuka. |
| Role bundles ke tor par Cowork plugins (Legal, Finance, Sales) | Desktop course ke plugins skills, connectors, slash commands aur sub-agents ko role mein bundle karte hain. Poora plugin system web se dastiyab hai ya nahin weeks-old detail hai jo live pages par check honi chahiye, magar desktop is ka asal ghar hai. |
| Backup-based recovery (Time Machine, git, Drive history) | Desktop course: "local backup file wapas lata hai." Web par local file nahin. Recovery Concept 5 discipline hai: pichla version tier 3 mein rakhein. |
Imaandaar khulasa. Cowork course aur yeh course aik kamray ke do darwazay hain. Agar kaam documents aur connectors hai aur lowest-friction shuruat chahiye, web aaj us course ka zyada hissa, Skills system samet, deti hai. Jab kaam apni machine ki files chhooay, data vendor servers se door rahe, ya OpenWork ka own-model path chahiye, desktop zaroori hai. Yeh web ki limit nahin. Desktop ke agla course honay ki wajah hai: kaam machine par ghar aaye to aap wahan jate hain.
Appendix C · Loop Engineering, web par
Loop Engineering course aisa system banana sikhata hai jo aap ki taraf se agents ko prompt karta hai: khud shuru, kaam, result check, kiya hua yaad, aur sirf insaani chunao par aap ko bulata hai. Yeh Claude Code aur OpenCode mein sikhata hai. Is course ke reading lens wali bilkul wohi six-part shape bhi istemal karta hai: heartbeat, chaar working parts aur spine.
Imaandaar taqseem yeh hai. Woh course khud lakeer kheenchta hai. Poora section "loop kahan chal sakta hai?" aur seedha jawab: plain chat box loop nahin chala sakta, Claude Code, OpenCode aur Cowork chala sakte hain. Web darmiyan mein hai. Saaday heartbeats fire kar sakta hai. Real loop ko trustworthy banane wali engineering nahin kar sakta.
Saadah shakal mein transfer:
| Woh course sikhata hai | Web par saada version |
|---|---|
| Scheduled heartbeat (har subah, laptop off) | Concept 9 ke scheduled tasks. Web yeh heartbeat achi karti hai. Clock par report bilkul Cowork ke /schedule aur ChatGPT Work Scheduled Tasks ka kaam hai. |
| Event heartbeat (kuch aane par react) | Kuch had tak. ChatGPT Work tabdeeli dekh sakta hai, event trigger ki halki shakal. Course ki GitHub-event aur API-trigger machinery nahin. |
| Spine (runs ke darmiyan memory) | Kamzor shakal. Account sessions aur Projects rakhta hai, kuch memory chalti hai. Magar repo ka progress.md aur rules file wala asal spine nahin jo har run pehle parh aur aakhir update kare. |
| Skills (aik baar likhi knowledge, har run parhi) | Eligible plans par dono families packaged Skills rakhti hain. "Koi run zero se nahin" hissa web par dono taraf transfer hota hai. |
| Heartbeat menu (task end, repeat ya aik baar?) | Concept 9 ka check-yourself. Halka loop chunne ka discipline surface se azad soch hai, command nahin. |
Transfer nahin hota, aur course asal mein isi par hai:
| Woh course sikhata hai | Web kyun nahin kar sakta |
|---|---|
| Maker-checker split (aik agent likhe, alag agent grade kare) | Course ka sab se aham khayal, web real version nahin kar sakti. Vendor ka run-until-done milta hai jahan vendor "done" tai karta hai. Aap apne model aur written rubric wala alag checker agent author nahin karte. Appendix A ka verification loop isi deewar par aata hai. |
| Worktrees (isolation taa ke parallel agents na takrayein) | Worktree machine par alag git checkout hai. Web par local repo nahin, isolate karne ko kuch nahin. Poora concept us local filesystem ko assume karta hai jo web nahin rakhti. |
Repo mein spine (progress.md aur rules file) | Asal spine apni repository ki files hai, har beat ke shuru mein read aur aakhir mein update. Web memory vendor ke format mein hai. Course ka "no spine, no loop" rule batata hai sanjeeda loops ko repo-owning tool kyun chahiye. |
/goal, /loop, capped shell loops (aap ka controlled run-until-done) | Claude Code commands aur OpenCode shell patterns. Web vendor version deti hai, course wala nahin jahan stopping condition, retry cap aur no-progress check aap set karte hain. |
| Dynamic workflows (poori orchestration re-runnable script) | Loop body ki gehri shakal: kai subagents, checked aur capped, aik script mein. Local tools mein bhi sirf Claude Code. Web se bohat aagay. |
| Cloud Routines aur GitHub Actions (unattended runtime) | Asal heartbeat Anthropic Routines ya GitHub runners par repo, branch rules aur run caps ke saath. Web scheduled task halka cousin: report karta, reviewed pull request claude/ branch par nahin kholta. |
Imaandaar khulasa. Web loop ke sab se saaday heartbeat ki shape chala sakti hai: sotay waqt scheduled report. Yeh real aur bohat logon ke liye kafi. Magar Loop Engineering bunyadi tor par heartbeat nahin. Un parts par hai jo unattended loop ko safe banate hain: alag checker, isolation, apna spine, controlled stopping conditions. Har cheez repo own aur code run karne wala tool mangti hai. Yeh appendix saaf promotion dikhata hai. Jab scheduled report ko aisa loop banana ho jo sotay waqt cheezein badlay aur aap bharosa karein, web se aagay barh chuke hain, aur Loop Engineering wahan milta hai.
Appendix D · Harness Engineering, web par
Poora course isi appendix ki taraf ishara karta raha. Concept 3 aur harness-war note mein taqseem batayi thi. Web par finished harness rent karte hain. Baad ke courses banana sikhate hain. Harness Engineering course mein woh building hoti hai. Yeh model ke gird woh layer sikhata hai jo raw intelligence ko trustworthy agent banati hai: agent kya kar sakta, kya jaanta, kaam kaisay prove hota, aur ghalat hone par kya hota hai.
Is liye yahan sab se tez lakeer hai. Building ka taqreeban kuch web par transfer nahin. Magar course ki soch, paanch verbs, mukammal transfer hoti hai. Web user ke liye behtareen lens: rented harness kya karta aur kya nahin.
Lens ke tor par paanch verbs (soch mukammal transfer):
Course harness ke har kaam ko paanch verbs mein rakhta hai. Web par aap koi nahin banate. Magar sab paanch dekh sakte hain, aur inhein jaan kar samajhte hain vendor se kya karwane par trust hai.
| Verb | Matlab | Web par kaun karta hai |
|---|---|---|
| Constrain (agent ki had) | Permission rules, deny lists, sandboxes | Vendor, aap ke liye. Connector scopes aur approval mode (Concepts 6 aur 7) chota controlled hissa hain. |
| Inform (zaroori maloomat) | Rules file, skills, tool design | Kuch aap. Project instructions aur packaged Skills (eligible plans par dono products) aap ki inform surface. Baqi vendor ka. |
| Verify (kaam ko count se pehle prove) | Hooks, tests, typed output | Vendor ka built-in run-until-done. Apna hook ya check nahin shamil kar sakte. |
| Correct (recover, phir system fix) | Checkpoints, ratchet | Vendor ka. Wapsi ka sirf Concept 5 ka tier-3 copy. |
| Escalate (insaan ko bhejna) | Human gate, loud logs | Shared. Concept 7 ka phone gate aap ki escalation surface. Logs vendor ke. |
Table imaandaar tasveer hai. Paanch mein se do verbs mein web par haqeeqi hissa: Projects aur packaged Skills se inform, phone se escalation gate par baithna. Baqi teen, jo unattended agent ko safe banate hain, aap ke liye aise tareeqon se hote hain jinhein dekh ya badal nahin sakte.
Transfer nahin, kyun ke harness banana poora course hai:
| Woh course sikhata hai | Web kyun nahin kar sakta |
|---|---|
| Permission rules (blast radius se allow, ask, deny) | Deny rule settings.json ya opencode.json ki enforced line hai. Web approval mode deti hai, rule language nahin. Web par ".env parhna deny" nahin likh sakte. |
| Core rule: guardrail harness mein, prompt mein kabhi nahin | Course ka sab se aham jumla aur web par jis par action nahin le sakte. Rule ko prompt se enforced layer mein le jane ke liye layer ki access chahiye. Web expose nahin karti. Web "guardrails" vendor ke, plus Project instructions ki baat, jise course request kahega, guardrail nahin. |
| Sandboxes (filesystem, network, branch fences) | Sandbox agent ki rasai ke gird OS-level deewar. Aap ki controlled machine aur fenceable runtime assume karta hai. Web session vendor ke servers aur sandbox mein, un terms par jo aap set nahin karte. |
| Hooks (action se pehle/baad automatic code) | Hook aap ka code hai jo harness fixed lamhay par chala kar action block/check karta hai. Web aap ka code nahin chalati. Appendix A ki wohi deewar, harness taraf se: sanjeeda control ko local tool chahiye. |
| Typed output (machine-checkable validated verdict) | Aap validator likhte hain jo loop result trust karne se pehle chalata hai. Controlled runtime chahiye. Web vendor verification deti hai, apni shamil karne ki jagah nahin. |
| Ratchet aur chaar failure classes (har ghalati permanent fix) | Ratchet har sabaq aap ki owned surface mein likhta hai: rule, hook, fence. Web par likhne ki surface nahin. Failure classify karne ki aadat seekh sakte hain, fix us jagah install nahin. |
| Inner/outer split, observability, harness coupling | Waqt ke saath harness banane/tune karne wale engineer ke disciplines. Assume karte hain harness aap ka hai. Web user design se own nahin karta. Rent karte hain, tune nahin. |
Imaandaar khulasa. Harness Engineering section ka sab se saaf "browser chorein" course hai, aur Concept 3 se yeh course isi ki taraf ishara kar raha. Is ka poora mazmoon woh layer hai jo web par rent aur desktop par build hoti hai. Aaj aik cheez paar le jayein: paanch verbs, rented harness ko dekhne ka lens, aur rule ke real guardrail enforce hoti hai, request nahin. Jis lamhay guardrail set, deny rule, hook, sandbox fence ya fix ratchet karna ho, woh exact edge hai. Web par harness trust karte hain. Is course mein banate hain. Box rent aur engineer karne ka farq.
Appendix E · Trusting the Checker, web par
Trusting the Checker course woh sawal poochta hai jo loop aur harness courses ne roka: kaise jaanein checker acha hai? Jab agent apna kaam grade ya reviewer PASS de, verdict model ki rai hai. Yeh course evals sikhata hai: tester ko test karna, taa ke "checker ne PASS kaha" defendable number ho. Golden set of case files banata, headlessly chalata, judge ko apni grading se calibrate aur har change result par gate karta hai.
Yeh section ka sab se gehra "repo aur shell chahiye" course hai. Poora method files ka folder, runner script aur jq hai. Machinery web par nahin. Magar measurement ko dekhne ka tareeqa har agent output, web agent samet, ko badalta aur mukammal transfer hota hai.
Judgment ke tor par transfer (web par asal value):
| Woh course sikhata hai | Web par zehni aadat |
|---|---|
| Agent distribution hai, function nahin (run nahin rate grade) | Sab se mufeed khayal. Scheduled report aik baar achi chale to aik run, proof nahin. Aham ho to trust se pehle kai runs check karein. Tools nahin, aik green result par trust na karne ka discipline. |
| Judge bhi model hai (PASS calibration mangta hai) | Web agent "done" ya "check kiya" kahe to claim, proof nahin. Vendor checker calibrate nahin kar sakte. Khud human calibration rahein: khaas tor par shuru mein verdicts spot-check. |
| Kitne se pehle kaun se fail parhein (rate severity chhupati hai) | Batch review mein ghalat teen items high overall count se zyada aham. Har surface ke output par aadat kaam karti hai. |
| Bar decision hai, discovery nahin | Kitna acha kafi aur kin tasks ke liye, director ke tor par aap ka faisla. Har surface par judgment aap ka. Concept 8 ka plan review results par. |
| Demos sab se kamzor evidence | "Mere try mein chala" aik run, kaam jo chalne ki wajah se chuna, umeed rakhne wale ke samne. Aham web task par trust mein yeh ehtiyat bachati hai. |
Transfer nahin, kyun ke eval suite repo aur shell mangti hai:
| Woh course sikhata hai | Web kyun nahin kar sakta |
|---|---|
| Golden set (version-controlled case files folder) | Cases repo ki JSON files, code jaisi reviewed aur dated. Web par repo ya version karne ki jagah nahin. |
Runner (claude -p / opencode run, shell loop + jq) | Poora runner headless command-line execution aur shell grading. Web na shell na scripts chalati hai. |
| Judge calibration (blind 20-item protocol, four-cell table) | Soch haath se ho sakti, magar course inspect/swap honay wale model judge ko measure karta hai. Web judge sealed vendor ka, protocol nahin chala sakte. |
| Regression suite aur CI gate (har change par rerun, baseline se neeche fail) | Branch protection aur committed baseline ke saath CI eval job. CI, branches, baselines repo tools hain. |
| Schedule par drift watching (nightly runs, drop alert) | Poora suite baseline ke khilaf scheduled loop. Web report schedule, committed baseline ke khilaf suite nahin, kyun ke suite/baseline rakhne ki jagah nahin. |
| Goodhart's law defenses (sealed hold-outs) | Agent ki rasai se bahar folder mein cases. Aap ka controlled aur fenced filesystem assume karta hai. Web na folder na deewar deti hai. |
Imaandaar khulasa. Trusting the Checker web se sab se kam buildable, kyun ke poora method files, shell aur repo hai. Web par eval suite nahin bana sakte. Magar jaldi parhne ke liye shayad sab se qeemti, kyun ke soch ko tools nahin chahiye. Har web agent par bharosa karte waqt aap untested checker trust karte hain. Course is ka wazan, run ke bajaye rate grade, volume se pehle severity aur human calibration rehna sikhata hai. Repo par asal suite banayenge. Tab tak mindset: imaandari se dekha track record hi trust kamata hai, aur web par record aap rakhte hain.
Appendix F · Leaving the Laptop, web par
Yeh aakhri course-mapping appendix aur poore course ki boundary se loop band karta hai. Shuru se yeh course khud ko front door aur Leaving the Laptop ko floor plan kehta raha. Ab wajah nazar aati hai. Woh section ka aakhri sawal hal karta hai: agent kya nahin, kahan rehta aur kaun zinda rakhta. Har option chaar homes mein, aur aaj ki web surface map se bahar nahin. Khaas limits wala khaas home hai.
Is liye yeh pichlay paanch se mukhtalif hai. Woh poochte thay: us course ka kya web par? Yeh poochta hai: us course ke map mein web kahan? Jawab section mein is course ki jagah saaf karta hai.
Chaar homes mein web surface:
Leaving the Laptop aik sawal se runtimes sort karta hai: loop ka operator kaun, kaam kahan? Web yahan utarti hai.
| Us course ka home | Web surface ki jagah |
|---|---|
| Home 1: aap ka session (sab own; lid band par mar jata) | Haath se chalaye plain web session ka qareebi match. Aap shuru/dekhte, khatam par rukta. Web "lid band par marta" bhi hata deti hai, session company servers par hai. |
| Home 2: cloud schedule (clock bahar; laptop off) | Web ki had, aur achi pohanchti hai. Cowork ya ChatGPT Work Scheduled Task knowledge work ka bilkul yeh home. Course Cowork Scheduled Tasks ko home-2 example kehta hai. Concept 9 ki report home-2 loop hai. |
| Home 3: managed runtime (definition bhejein; vendor loop operate) | Author ke tor par web se bahar. API se agent definition vendor ko aur business operate karna builder ka move, SDK/platform docs se, browser chat se nahin. |
| Home 4: aap ka process (aap ke servers par harness library) | Mukammal bahar. Mode 2, runtime khud likhein/chalayein. Kitab mein web se sab se door. |
Deciding judgment ke tor par transfer:
Course ka tohfa mechanics nahin. Home chunne ke chaar sawal, pure judgment, har surface par.
| Woh course sikhata hai | Web par bhi poochein |
|---|---|
| Q1: user kaun? (sirf aap ya doosray?) | Sab se aham. Sirf aap ho to web scheduled task aksar poora jawab; course kehta hai home 2 aksar had. Doosray output par depend karein to edges mehsoos hoti hain. |
| Q4: buri raat ki qeemat? (blast radius speed tai) | Concept 9 ki report-versus-act aur harness budgeting, runtimes par. Har surface par judgment aap ka. |
| Headless pul hai | Web se nahin guzarte, aur yehi nishani. Web se mazboot home ki har move headless command (claude -p, opencode run) se, jo web nahin deti. Pul browser ke aakhir se shuru. |
| Suitcase test (discipline safar, mechanics rebuild) | Standing briefs, Project instructions, "done" ki samajh discipline hai aur mazboot tool mein saath. Buttons/settings nahin. Working knowledge apne notes mein portable rakhna "repo sach rakhta hai" ka web version. |
Transfer nahin, kyun ke laptop chorna browser chorna hai:
| Woh course sikhata hai | Web kyun nahin kar sakta |
|---|---|
Headless invocation (claude -p, opencode run) | Har move ka pul. Web ka headless command nahin, isi liye homes 3/4 nahin, home 2 aadha. |
| Managed runtime (API par definitions, environments, sessions) | Code aur platform docs se chalaya builder runtime. Browser path nahin. |
| Minimum unattended kit (idempotency, missed-run detection, concurrency locks) | Operated loop ke gird six controls. Web par vendor loop operate karta, jo de woh inherit, izafa nahin. |
| Arrival protocol aur re-baselining (eval suite se move prove) | Naye home mein golden set, Appendix E ki suite aur repo chahiye. Suite nahin, protocol nahin. |
| Runtime own karna (homes 3/4) | Course ka pichla aadha. Infrastructure operate karna web rent ka ulat. |
Imaandaar khulasa aur course-mapping appendices ka close. Leaving the Laptop batata hai web ki haqeeqi jagah aur had. Haath se kaam ke liye acha home 1, schedule par knowledge-work reports ka haqeeqi home 2. Bohat real work aur kai logon ke liye poori kahani. Magar headless bridge, repo, own sandbox ya controlled runtime se doosron ki service par edge milti hai. Is course ne pehlay page se yeh map ki. Yeh course front door, Leaving the Laptop poore ghar ka floor plan. Ab jaante hain browser se kaun se kamray khulte aur kin ke liye keys/install chahiye. Web surface ki imaandaar shape: real shuruat, saaf aakhir, aur aagay sab ka map.
Appendix G · Harness war, poori industry mein
Ooper chhay appendices ne is course ko baad ke har course se map kiya. Aakhri bahar dekhta hai. Concept 3 ne do companies ki harness war kahi. 2026 tak do nahin, poori industry, aur taqreeban har sanjeeda lab ki framing aik: larai raw model se hat kar us ke gird harness par.
Yahan sab se mufeed cheez yehi dekhna hai. News headlines models par: yeh score zyada, woh sasta. Magar products wrapper par muqabla: schedule, connectors, memory, gate, runtime. Yeh field guide hai kaun kahan lar raha. Yaad karna maqsad nahin. Products hafton mein purane. Six-part shape har aik ko parhti hai.
Industry snapshot, facts checked: 21 July 2026. Yeh appendix sab se jaldi purana aur akelay refresh ke liye hai. Kai items announced, shipped nahin, tasveer mahana badalti hai. Dated snapshot samjhein, lasting fact nahin, bharosay se pehle sources se har row verify.
Players teen battlefronts par. Appendices ki rent-versus-build lakeer poori industry mein.
Front 1: consumer work surface (finished harness rent). Isi course ki larai: chat box ke paas agent surface, rozmarrah knowledge work.
| Player | Surface | Haalat |
|---|---|---|
| OpenAI · ChatGPT Work | ChatGPT ke andar mode | Course ka mazmoon. July 2026 shipped. |
| Anthropic · Claude Cowork | claude.ai par remote session | Course ka mazmoon. Web/mobile July 2026 expanded. |
| Google · Gemini Spark | Gemini app mein 24/7 cloud agent | I/O 2026 announced: device off hone par Google cloud machines par Gmail, Docs aur Workspace mein personal agent. Wohi shape, wohi discipline: paisay kharch/email se pehle poochta. Mid-2026 sirf early US access. Search, Workspace, Chrome se gehri jeb aur wasee distribution. |
| Microsoft · Copilot Tasks | Microsoft 365 mein unified tasks surface | Auto, Researcher, Analyst modes, scheduling: wohi parts enterprise Office work par. Faida distribution. Analyst: larai best chatbot ki nahin, work surface own karne ki. |
| Meta · Meta AI | WhatsApp, Messenger, Instagram mein agents | Capability nahin reach par: harness jahan billions chat karte; Muse Spark rebuild ka signal. Agent startup Manus ki reported two billion dollars deal Chinese regulators ke reverse order par 2026 mein khatam, geopolitical front ki yaad. |
Concept 3 ke chhay parts se table parhein. Har aik: heartbeat, connectors, run-until-done loop, state spine, human gate, body. Names/owners alag, shape aik. Paanch companies mein aik saath thesis.
Front 2: managed runtime (surface nahin, runtime rent). Consumer se aik layer neeche enterprise platform: vendor ko agent definition, vendor loop operate. Appendix F ka home 3 product ke tor par.
| Player | Platform | Kya hai |
|---|---|---|
| Google Cloud · Gemini Enterprise Agent Platform | Vertex AI ka evolution | Agents build/run/govern, orchestration, runtime, memory aur Agent Identity/Gateway jaisay governance primitives. Enterprise roadmap is brand se. |
| Microsoft · Copilot / Agent 365 | Microsoft 365 ka enterprise agent layer | Wohi governance push, Office, Teams, Windows se jura. |
| Alibaba Cloud · Wukong, Qwen Cloud | AI-native enterprise platform | Multiple agents se business tasks automate, Skills portal 60-plus cloud products ko MCP-compatible skills banata. |
Front 2 mein through-line governance, capability nahin. Mushkil kaam working agent nahin, enterprise ke liye trustworthy: identity, policy, audit. Survey mein aksar firms production agents, bohat kam govern kar sakti hain. Yehi Harness Engineering aur Trusting the Checker ka gap, enterprise product ke tor par.
Front 3: open engine (apna harness build). Rented consumer surface se door open-weight model: surface nahin, sirf engine jo aap apne built/run harness mein. Rent-versus-build ka build end.
| Player | Engine | Yahan kyun aham |
|---|---|---|
| DeepSeek · V4 | Open-weight, tool-native, Anthropic-compatible API | Claude Code, OpenCode aur coding agents mein drop-in pre-tuned. Adoption real: release ke mahine mein fast model DeepSeek agentic traffic ka zyada hissa, Chinese open models mid-2026 token share mein American se aagay. |
| Alibaba · Qwen 3.6 / 3.7 | Open-weight (Apache 2.0), plus proprietary "Agent Frontier" flagship | Native MCP, 1-million-token context, 35-hour autonomous run aur 1,000-plus tool calls (proprietary Agent Frontier ka run, Apache open line nahin). Open line self-hosting pick. |
| Meta · Llama | Open-weight | Ab bhi mashhoor open line, agarche Meta strategy pure model releases se agents (Manus) khareed/release ki taraf. |
Front 3 mein kitab ka gehra wada hai. Open-weight engine se hi poora harness own: model, runtime, sandbox, data. Wajah Concept 10: hosted DeepSeek API China mein, regulated/sensitive data ke liye open weights self-host clean path. Engine own, custody own. Ghar ki door deewar jis ka front door course ne khola.
Imaandaar khulasa. Teen fronts, aik war. Consumer par paanch companies best finished harness rent karne ki race. Managed runtime par giants enterprise ko harness rent aur governance par muqabla. Open engine par labs apna banane ke parts. Models, prices, distribution alag. Shape nahin. Teeno fronts ka har product heartbeat, connectors, loop, spine, gate, body. Gentle surface par aik baar shape seekhein aur har saal kisi front ki industry parh sakein. Isi liye kitab shape sikhati hai, tool nahin. Tools war mein. Shape woh peace treaty hai jo darmiyan le jate hain.
Sources aur mazeed mutala
Is kitab ke andar
- Skills & Connectors, do upgrades jinhein yeh course browser se chalata hai. Foundations Course 5.
- Cowork & OpenWork, is course ke delegation loop ki desktop gehrai: folders, trust levers, autonomy ladder, regulated workloads.
- Loop Engineering, chhay parts jinhein course ne reading lens ke tor par liya, design ki jane wali cheezon ke tor par. Aur action schedules ka safe version.
- Leaving the Laptop, is surface ke peeche runtime decision. Yeh course front door, woh floor plan.
Products (live pages, sirf yahin current numbers)
ChatGPT Work (OpenAI):
- Launch announcement, availability aur examples: ChatGPT ab aap ke sab se ambitious kaam ka partner hai (9 July 2026).
- ChatGPT Work aur Codex, Work/Codex ka rishta, usage structure, desktop vs web/mobile.
- ChatGPT mein Scheduled Tasks, cadence types (once, schedule, event, monitor), aur Project-file caveat.
- ChatGPT aur Codex mein Plugins, connector layer jisay yeh course connectors kehta hai.
- ChatGPT mein Skills, packaged Skills, creation aur eligible plans.
Cowork (Anthropic):
- Web, desktop aur mobile par Claude Cowork istemal karein, remote sessions, per-surface features aur local files ke desktop-bridge rules.
- Claude Cowork architecture overview, remote/local sessions ki isolation aur har aik ki rasai.
- Claude Cowork mein recurring tasks schedule karein, cadence aur kisi device ke online hue baghair remote execution.
- Claude Cowork mehfooz istemal karein, Concepts 6 aur 7 ke peeche trust model.
- Claude Code (alag coding surface) aur repo-attached cloud runs: code.claude.com/docs.
OpenWork (open-source Cowork twin):
- OpenWork, open-source Cowork alternative (OpenCode se powered), aur docs.
Wasee industry (Appendix G), primary aur launch sources:
- Google: Gemini Enterprise Agent Platform (Google Cloud, I/O 2026); I/O 2026 ka consumer agent Gemini Spark, 24/7 cloud agent (mid-2026 tak early US access).
- Microsoft: Auto, Researcher aur Analyst modes ke saath Copilot Tasks (Microsoft 365 Copilot); "sirf wohi dekhe jo aap dekh sakte" permission model ke liye Researcher agent.
- Meta: Manus acquisition ko Chinese regulators ne unwind karne ka hukum diya aur Meta 2026 mein reverse karne laga; Manus-based strategy claim ko unsettled samjhein.
- DeepSeek: V4 agentic adoption (OpenRouter) aur V4 coding-agent integration guides.
- Alibaba: Qwen3.7, the Agent Frontier aur Qwen-Agent framework (open, MCP-native).
Kisi number par bharosa karne se pehle in pages se verify karein. Yeh kitab ka sab se jaldi aging page hai: dono web work surfaces July 2026 mein launch ya khaas tor par expand hui (Cowork web/mobile ne mojoodah product barhaya, ChatGPT Work naya mode), rollouts plan ke mutabiq staged, allowances metered, names badalte hain. Har cheez 21 July 2026 ko check hui. Volatile facts (rollout tiers, allowances, connector actions, Project-file caveat) ke saath wahan source marker bhi hai jahan nazar aate hain.
Aik-line khulasa
Chat box intezar karta hai aur hamesha karega. Bagal ki surface badli: aisa kaam jo aap ke jane ke baad company ke servers par chalta aur zaroorat par phone se aap tak aata hai.
Sari aadatein aik jagah
Poora brief dein. Maslay plan par pakrein, baad mein nahin. Finished work exit (Tier 3) mein bhejein. Sirf reporting work schedule karein, aur sirf woh jis se pehle door ja chuke. Har naya product chhay parts se parhein. Aur is sab par bharosa karne se pehle runtime sawal Leaving the Laptop mein le jayein.