Laptop Chhorna: Runtime Crash Course

12 Concepts · Aise loop se jo lid band hote hi mar jata hai, apne home wale worker tak

Aapka system tayyar hai, aur phansa hua hai. Aapka design kiya loop har subah 9 baje chalta hai. Harness khatarnaak actions rokta hai. Eval suite reviewer ko aisa number deti hai jis ka aap difa kar sakein. Aapne sab kuch sahi kiya, lekin laptop band karte, Wi-Fi chali jate, ya flight mein baithte hi sab ruk jata hai. Aapke banaye sab se bharosemand worker ka aik single point of failure baqi hai: desk par rakhi machine.

Yeh course section ke aakhri khule sawal ka jawab deta hai: runtime decision. Agent kya karta hai, yeh nahin, kyunke teen courses woh tay kar chuke. Sawal yeh hai ke woh kahan rehta hai aur use zinda kaun rakhta hai: jagah aur dekh bhaal karne wala, behavior nahin. Yeh infrastructure ki detail lagti hai. Hai nahin. Yehi sawal tay karta hai ke aap bahut ache tool ke operator hain ya aise worker ke owner jo aapke aaye baghair bhi aa jata hai. Company scale par pehla Digital FTE banate waqt bhi yehi sawal dobara poochhenge. Yahan itne chote scale par seekhein ke tamam parts nazar aayen.

Shuru karne se pehle aik wada. Yeh course aapko DevOps engineer nahin banata. Har move wohi cheezein istemaal karti hai jo aapke paas hain: agentic coding course ki config files, evals course ki headless commands, loop course ke schedules, aur move ke kaam karne ka saboot dene wali suite. Sirf aik waqai nai cheez, yani aisa runtime jo koi aur aapke liye operate kare, seedhi tarah introduce hoti hai, price tag aur trade-offs saamne rakh kar.

Pehle yeh chahiye: Checker par Bharosa. Us course ne eval suite di, aur yeh course us par gehra depend karta hai: naye runtime ki move bilkul woh change hai jis par bharosa karne se pehle suite dobara chalni chahiye. Yeh poori Stage 3 trilogy assume karta hai: Loop Engineering (beats, spine, human gate), Harness Engineering (paanch verbs, ratchet), aur evals course (golden set, baselines, drift). Agar yeh alfaaz naye hain to pehle woh courses karein. Yeh course unki banai machinery move karta hai.

Yahan naye hain? Jo pehle aana chahiye uska 2-minute recap

• Beat: scheduled loop ki aik poori run. Morning-triage loop har weekday aik beat chalata hai.
• Harness: woh layer jo tay karti hai ke agent kya kar sakta hai, use kya jaanna chahiye, kaam ka saboot kaise mile, aur ghalat hone par kya ho.
• Golden set: real pakri failures se bana eval cases ka folder, jo har change par dobara chalta hai.
• Baseline: recorded pass rate jis se nai runs compare hoti hain. Baseline se drop alarm hai.
• Drift: aapke change ke baghair behavior badalna, aam tor par neeche ka model update hone se.
• Human gate: risky ya failed work kisi insan ke paas jata hai. Koi unattended cheez main tak nahin pahunchti.
• Headless mode: agent ko interactive session ke baghair chalana (claude -p Claude Code mein, opencode run OpenCode mein), taake script ya schedule use drive kar sake.

Agar in mein se koi cheez nai hai to pehle Stage 3 ke teen courses parhein. Yeh course unki banai aur prove ki machinery relocate karta hai.

Aam alfaaz mein aham istilahaat

Istilah	Aam matlab
Runtime	Computer aur agent ke gird software jo asal mein use chalata hai: shuru karta, input deta, aur crash par restart karta hai.
Home	Runtime choice ke liye is course ka lafz: aapka session, cloud schedule, managed runtime, ya aapka apna process.
Control plane	Runtime ka hissa jo agent loop operate karta hai: sessions, scheduling, event streams, restarts.
Execution plane	Jahan kaam land karta hai: tools ka sandbox aur unka chhua data. Dono planes ke owners alag ho sakte hain.
Custody	Data kis ke paas aur control mein hai: kin machines par rehta aur kaun use dekh sakta hai.
Headless	Agent ko interactive session ke baghair command se chalana. Har move ka bridge.
Schedule	Aapke baghair muqarrar waqt par loop shuru karne wali cheez: Claude Code Routine, Cowork Scheduled Task, ya scheduled GitHub Actions job.
Managed runtime	Service jahan aap agent definition bhejte hain aur vendor control plane operate karta hai; default mein sandbox bhi, lekin zaroori nahin.
Agent definition	Agent ko chalaye baghair describe karne wali har cheez: model, system prompt, tools, rules, guardrails.
Session (hosted)	Managed runtime ke andar aik ongoing kaam, apne preserved state aur event log ke saath.
Sandbox	Band environment jahan agent ke actions execute hote hain. Kuch homes mein model ke paas, doosron mein alag.
Portability	Naye home ki move par bina dobara banaye system ka kitna hissa bach jata hai.
Lock-in	Home chhorne ki cost, har us cheez mein jo dobara banana pare. Kam portability, zyada lock-in.
Blast radius	Kisi home ki buri raat kahan tak pahunch sakti hai. Harness course ka budgeting sawal, runtimes par.
Pager	Raat ko system tootne par engineer ko jagane wali alarm. Pager carry karna uski zimmedari lena hai.
Re-baseline	New home ke existing acceptance bars pass karne ke baad runtime-specific pass rate record karna, old baseline ko comparison aur history ke liye bachaye rakhna.
Probation	Aazmaishi muddat: poora bharosa karne se pehle new home ko qareeb se dekha jata aur old home available rehta hai.

Yeh kahan se aya

Industry ki zyada tar tareekh mein "deployment" developer ka lafz tha: code likha, phir servers par ship kiya. 2025 aur 2026 mein yeh badla ke operators, yani jin logon ne agent likha nahin balkay configure aur prove kiya, unke paas achanak deploy karne layak cheez aa gayi. Measured track record wala configured loop asset hai, aur sirf laptop khula hone par maujood asset buri tarah stored hai. Vendors ne yeh dekha. Anthropic ne consumer tools ke liye cloud execution, phir agent definitions ka hosted runtime ship kiya. Open-source side ne low-cost option ke liye CI schedulers standard bana liye. Is liye runtime decision, jo pehle project ke aakhir mein sirf engineers ka sawal tha, ab bahut jaldi aata hai aur un logon tak pahunchta hai jo kabhi server nahin likhenge. Yeh course unke liye hai. (Sources aakhir mein.)

Aik tasveer mein mindset shift

Yeh course do tools saath sikhata hai, jaise poori trilogy ne kiya. Judgment, yani kaunsa home, move kab aur uska proof kaise, dono mein aik hai. Mechanics pehle ke kisi course se zyada alag hain, aur course wajah saaf kehta hai: aik tool ka vendor uske liye cloud operate karta hai, doosre tool ka "vendor" aap hain. Yeh asymmetry footnote nahin. Runtime decision isi ke baare mein bhi hai.

caution

Mid-July 2026 mein sach. Is course ki mechanical layer ki har cheez, product names, endpoints, prices aur flags, section ke kisi aur course se tez badalti hai. Move se pehle claude update ya opencode upgrade chalayein aur naam ya number par bharosa karne se pehle live docs (code.claude.com/docs, docs.claude.com, opencode.ai/docs) check karein.

Is course mein kya hai

Part	Mauzu	Aap kya seekhenge
1	Aakhri dependency	Laptop par proven loop buri tarah stored asset kyun hai, aur har option sort karne wala aik sawal
2	Headless bridge hai	Har home ki shared invocation aur pehli move: cloud schedule
3	Managed runtime	Agent definitions, environments aur sessions: kya dete hain, kya milta hai, cost kya hai
4	Move khud	Kya travel karta, kya dobara banta, aur new home mein track record dobara kyun kamana hai
5	Home chunna	Chaar sawal, homes ko jaan boojh kar mix karna, aur aik poori end-to-end move
6	Imaandar rehna	Lock-in, ownership drift, jo koi home fix nahin kar sakta, aur Mode 2 ka bridge
Live	Dogfooding	Is book ke loops kahan rehte aur kyun
Practice	Projects	Aasan se mushkil aath moves

Kar ke seekhna chahte hain? Aik poori move dekhne ke liye pehle Part 5 parhein. Phir baqi parts par laut aayen.

Is course ko parhne ke do tareeqe

Pehli baar? Parts 1 se 5 order mein, "Going deeper" wali har note skip kar ke. Is reading mein taqreeban nabbe minutes hain. Phir Projects 1 se 3 karein, jo zyada waqt lenge. Unke baad loop aise schedule par chalega jise laptop nahin chahiye, aur aap number ke saath keh sakeinge ke woh ab bhi kaam karta hai.

Doosri reading (unattended raaton ke pehle maheene ke baad): Part 6 poora, deeper notes aur Projects 4 se 8. Lock-in tabhi poora samajh aata hai jab dobara banane wali koi qeemti cheez ho.

English second language ke taur par parh rahe hain? Course ki har picture-word ka aam version qareeb hai. "In simple terms" boxes aur oopar ka glossary wahi matlab chote seedhe jumlon mein dete hain. Jumla sajawati lage to qareeb box wahi baat seedhi kehta hai.

Kya yaad rakhna hai, aur kya dekhna hai

Do layers, mukhtalif raftaar se purani hoti hui. Pehli yaad rakhein. Doosri dekh kar lein.

• Dairpa layer. Loop kaun operate karta aur work kahan execute hota hai, yehi sawal hai; baqi detail. Headless har move ka bridge hai. Discipline travel karti, mechanics dobara banti, aur trust phir measure hota hai. Chaar sawal home chunte hain. Blast radius move ka waqt tay karta hai. Har home human gate rakhta hai.
• Mechanical layer. Neeche ka har product name, endpoint, price aur flag. Managed-runtime APIs course likhte waqt chand hafton purani hain. Har detail ko yaad karne ka fact nahin, live docs ki pointer samjhein.

---

📚 Teaching Aid

Poora slideshow kholein

Poora presentation dekhein: Laptop Chhorna: Runtime Crash Course

---

Part 1: Aakhri dependency

1. Laptop par proven loop buri tarah stored asset hai

Teen courses mein banai cheezein ginen. Aik loop jo aapke baithne se pehle morning queue triage karta hai. Aik harness jo khatarnaak ghaltiyan namumkin banata aur har pakri failure ko permanent rule banata hai. Aik eval suite jo beech ke reviewer ko defended number deti hai. Seedha kahein: written job description, supervisor aur performance record wala junior colleague.

Ab dependencies ginen. Laptop khula ho. Session logged in ho. Machine 9am par awake, power aur network par ho. Aik bhi miss ho to beat khamoshi se nahin chalti, aur loop course ne skipped beat ki cost sikhai: queue barhti, escalations jama hoti, aur Monday wala aap Friday ke band lid ki qeemat deta hai.

Isay aur teekhe tareeqe se kahein. Trilogy ne single points of failure aik aik kar ke hataye: maker-checker split ne single unreviewed opinion, harness ne single unguarded action aur evals ne single unchecked checker hataya. Aik baqi hai, aur woh aap hain: human gate ke paas sahi rehne wali judgment nahin, balkay aapka hardware. System apni machine se zyada reliable hai. Yehi signal hai ke woh apne home se bara ho gaya.

Aam alfaaz mein

Aapne acha worker train kiya, phir shart rakhi ke woh sirf aapke living room mein aur sirf aapki mojoodgi mein kaam kare. Worker theek hai. Arrangement masla hai.

2. Aik sawal har option sort karta hai: loop kaun operate karta, work kahan execute hota hai?

Naya home dhoondte hi options barhte aur vocabulary shor karti hai: cloud sessions, hosted agents, managed runtimes, SDKs, serverless yeh, orchestrated woh. Sab ko aik sawal se kaat dein. Iske do halves hain: agent loop kaun operate karta hai, aur uska work kahan execute hota hai?

Dono halves ke naam ab seekhna faida deta hai, kyunke modern runtimes unhein split kar sakte hain. Control plane loop khud hai: sessions shuru karta, model ko input deta, events stream karta aur 3am par crash restart karta hai. Execution plane jahan actions land karte hain: tools ka sandbox aur unka chhua data. Laptop par dono planes aik machine the, is liye farq zaroori nahin tha. Aage ke homes unhein alag kar sakte hain, aur sab se dilchasp home jaan boojh kar karta hai.

Har option un do halves se sorted chaar homes mein se aik hai, jaise oopar picture ne dikhaya:

• Home 1: aapka session. Config, runtime, uptime sab aapka. Trilogy yahin hui, aur building aur proof ke liye sahi home hai. Depend karne ke liye ghalat.
• Home 2: cloud schedule. Config ab bhi aapki: wahi rules file, skills aur subagents, lekin clock kisi aur ke computer par chali jati hai. Claude Code Routine aapka loop Anthropic cloud par chalati hai. Scheduled GitHub Actions job OpenCode loop GitHub runners par chalati hai. Sab se choti move, sab se bara foran faida.
• Home 3: managed runtime. Aap agent ki definition (model, prompt, tools, guardrails) dete hain aur vendor service control plane chalati hai: loop, sessions, crash recovery. Execution plane choice hai: default mein vendor cloud sandbox, ya custody ki zaroorat ho to aapki controlled infrastructure ka sandbox. Aap loop operate karna chhorte hain, gird ka business operate karte hain.
• Home 4: aapka process. Harness aapke likhe software ki library aur aapke servers ka hissa banta hai. Sab aapka, poori responsibility bhi. Yeh Agent SDK aur Mode 2 ka ilaqa hai; course darwaza dikhata hai, paar nahin karta.

Do planes ko seedha rakh kar dekhein. Home 3 asal mein do hai, is liye paanch rows:

Home	Control plane	Execution plane	Tootne par kaun jagta hai
1 · Aapka session	Aap	Aapka laptop	Aap
2 · Cloud schedule	Aap, scheduler ke zariye	Cloud runner	Shared
3 · Managed, cloud sandbox	Vendor	Vendor	Infrastructure unki, outcomes aapke
3 · Managed, self-hosted sandbox	Vendor	Aap	Plane ke mutabiq split
4 · Aapka process	Aap	Aap	Jaan boojh kar aapki

Ghaur karein sawal asal mein technical nahin. Har business har function par yehi poochta hai: khud karein ya kisi aur ko paise dein? Aap yeh soch pehle se jante hain. Baqi course wohi soch agent ke neeche computer par lagata hai.

Aam alfaaz mein

Chaar homes mein se chunna hai, seedhi nahin charhni: laptop (sab own karein), schedule (config own karein), managed runtime (definition own karein aur sandbox unke cloud ya custody par aapki infrastructure mein), ya apne servers (product scale par jaan boojh kar sab phir own karein). Rule: sirf woh own karein jo lazmi hai, jo dil chahe woh nahin. Zyada tar readers pehle teen mein rehte, aur aksar har loop ke liye alag homes saath chalate hain. Chautha home Mode 2 ka aaghaz hai.

Khud jaanchein

Lahore mein Ayesha freelance invoicing loop apne laptop par chalati hain. Load-shedding zyada tar shaamon power kaat deti hai, aur naya client roz local time 6pm par invoice baghair failure chahta hai. Unhein pehle kaunsa home chahiye, aur home 3 aaj unki zaroorat se zyada kyun hai?

Jawab dekhein

Home 2, cloud schedule. Masla sirf clock hai: loop proven, config kaam karti, lekin machine 6pm promise nahin kar sakti. Schedule ko cloud runner par move karne se power cut critical path se nikalta hai, aur new home ke siwa kuch dobara prove nahin karna. Client ke "without fail" par honesty note: CI schedules samet aksar schedulers taqreeban waqt promise karte hain, theek waqt nahin. Load mein delay aur kabhi scheduled run drop hoti hai. Is liye "without fail" scheduler feature nahin, Part 2 ki minimum unattended kit hai: missed-run detector, duplicate invoice na bhejne wali retry, aur 6:30 tak success record na ho to alarm. Home 3 doosre users, long jobs ka session state aur scale operations jaise masle solve aur unki cost leta hai, jo abhi nahin. Part 5 ke chaar sawal ise saaf test banayenge: user kaun hai? Aaj user Ayesha hain. Home 2 aur kit kaafi hain.

---

Part 2: Headless hi bridge hai

3. Har home headless samajhta hai

Is course ki har move mumkin banane wali khamosh baat yeh hai: evals course mein aap yeh bataye baghair bridge paar kar chuke hain ke woh bridge tha. Eval runner ne claude -p aur opencode run se agent ko guftagu nahin, command ki tarah call kiya. Koi khuli window nahin, sambhalne ko session nahin: prompt andar, kaam hota hai, output bahar, process band.

Yehi headless mode hai, aur poora course isi idea par khara hai. Jo cheez command chala sakti hai, woh ab aapka agent chala sakti hai. Shell script, cron job, CI runner, cloud scheduler, sab. Home 2 se aage har home andar se is sawal ka alag jawab hai: "headless command kis ka computer, kis ki clock par chalata hai?"

Evals course ki mechanics baghair badle saath aati hain: Claude Code mein machine-readable output ke liye --output-format json, OpenCode side par JSON event stream, aur reviewer ka verdict file mein taake koi prose parse na kare. Ab aik aadat jorein jab insan nahin dekh raha: headless run failure par zor se alarm de. Session mein error nazar aati hai. Runner par unchecked exit code aisi beat hai jo khamoshi se hui hi nahin, Concept 1 ki warning cloud mein dobara. Har headless wrapper exit code check karta aur failure par loud hota hai, yani loop course ke paanchveen verb wali saaf alarm uthata hai. Khamoshi ka matlab success hona chahiye. Yeh rule aap lagate hain, inherited default nahin.

Aam alfaaz mein

Interactive mode mein aap agent se baat karte hain. Headless mode mein koi script, schedule ya server agent ko note de kar result le leta hai. Agent note le sake to jahan note pohanchta hai, wahan kaam kar sakta hai.

4. Home 2, cloud schedule: pehle clock bahar jati hai

Pehli move jaan boojh kar sab se choti hai: banaya sab rakhein, sirf clock relocate karein. Config, yani rules file, skills, subagents aur hooks, bilkul wahi rehti hai. Sirf beat shuru karne wala badalta hai.

Vendor ka rasta wahi schedule hai jis ka loop course ne wada kiya tha, aur naam bhi wahi: Routine. Claude Code mein /schedule command (alias /routines) guftagu se banati hai. Iski configuration mein prompt, repositories, connectors aur environment save hote hain. Laptop khula ho ya nahin, woh Anthropic-managed cloud infrastructure par chalti hai. Repo-attached loop, "har weekday 9am par morning triage skill chalayein aur medium risk se oopar har cheez escalate karein," bilkul yeh ban jata hai.

Live docs se teen honesty notes, sab pehchane lagne chahiye:

• Routines research preview mein hain. Yeh mechanical layers mein sab se mechanical hai, depend karne se pehle verify karein.
• Design ke mutabiq scheduled run ghante se chand minutes baad shuru ho sakti hai (docs ise stagger kehti hain). Asal wada "taqreeban 9am" hai. Har scheduler ka yahi point vendor khud keh raha hai.
• Green run status sirf session ke infrastructure error ke baghair exit hone ka matlab hai. Iska matlab task success nahin. Concept 3 ka loud-failure rule prompt aur checks ke andar aapko lagana hai.

Aik trap ka naam friendly hai. Desktop app mein Local chunne se Desktop scheduled task banti hai jo aapki machine par chalti hai. Timer wala home 1, move bilkul nahin.

Claude Cowork mein knowledge work ke liye yehi idea Scheduled Task hai. Aik baar describe karein aur connectors, skills aur plugins ke saath schedule par remotely chalegi. Dekhein to results intezar karte hain. Wahi docs caveat deti hain: local files ya desktop applications wali task locally chalti hai aur laptop phir lazmi ban jata hai. Loop jo chhoota hai us se chunein: repository work Claude Code Routine, connector aur document work Cowork Scheduled Task. Dono mein laptop ka baqi kaam sirf results parhne ki jagah hai.

Do aur honesty notes, dono mechanical-layer. Pehla, cloud runner aapki machine nahin. Kin repos, connectors aur credentials tak pahunch hai, yeh configure hota hai, inherit nahin. Pehli Routine run is baat ki discovery bhi hai ke new home kya dekh sakta hai. Doosra, harness configured form mein travel karta hai, magar verify karein ke kiya. Permission walls aur hooks files hain, aur files sirf un runs ko bachati hain jo unhein load karein. Routines config kaise source karti hain, live docs se dekhein, aur Concept 8 ki probation non-negotiable rakhein.

Repository loops ke liye home 2 ka vendor-neutral version bhi hai jo loop course bana chuka: schedule: trigger wala GitHub Actions workflow. Repo config saath checkout karke CI runner par claude -p headlessly chalata hai. Wahi home, doosri company se rent. GitHub ka honesty note: scheduled workflows best-effort hain, load mein late, kabhi drop, aur sirf default branch se. Morning loop ka contract "taqreeban 9am" hai. Hard deadline ko neeche ki minimum unattended kit aur shayad mazboot scheduler chahiye.

OpenCode ka first-party hosted control plane nahin, aur course uska dikhawa nahin karta, kyunke imaandar version zyada sikhata hai. OpenCode mein home 2 woh scheduler hai jo aap chunte hain. Repo loops ke liye loop course ka GitHub Actions pattern use karein: opencode run ko schedule: trigger par chalayein aur .opencode/ config checkout karein. Baqi ke liye koi bhi cron machine: paanch-dollar cloud server, ghar ka always-on computer, ya scheduled runner.

Wohi milta hai jo tool ne pehle din kaha: aap aur loop ke darmiyan vendor nahin, model aur provider aapki pasand. Aap mante hain ke vendor ab aap hain: runner uptime, credentials aur updates aap sambhalein. Actions par repo loop mein burden chota aur zyada GitHub ka. Desk ke neeche chote computer par zyada kaam se home 1 dobara. Is liye Actions version is side ka default recommendation hai.

Tabs ka farq lesson hai, defect nahin: open tool poora runtime decision aapko deta aur kuch nahin chhupata. Yehi trade Personal Agent Harnesses section mein dobara milega, jahan runtime own karna hi point hai.

Jo version chunein, success ki definition aik aur measurable hai: laptop band rehne par kam az kam aik poora operating cycle aur das successful beats, khamoshi baseline ke ma'ni mein. "Cloud se aik baar chala" nahin; evals ne aik green run ki qeemat sikhai. Long-term uptime ka proof bhi nahin: initial operational evidence kahein, riskier ya rare loops ke liye barhayein. Scheduled beats baseline ke saath checked, alarm aik baar jaan boojh kar tested. Yehi home 2 mein move hai.

Minimum unattended kit. Course ne DevOps engineer na banane ka wada aik discipline ki jagah table se nibhaya. Insan na dekhe to yeh chhe controls optional nahin. Har aik chota aur uski kami ke peeche mashhoor failure story hai.

Control	Rule	Rokne wali failure
Idempotency	Retried beat dobara chalna safe ho	Invoice ya ticket do baar
Missed-run detection	Doosra system notice kare ke pehla shuru nahin hua	Na chala process apni absence report nahin kar sakta
Concurrency lock	Aik waqt mein aik beat, 8am zinda ho to 9am wait	Aik queue ki files par do agents
Credential discipline	Scoped service credentials, least privilege, rotation, personal login kabhi nahin	Cloud runner ke paas poori identity
Time semantics	Timezone likhi, daylight-saving aur catch-up pehle tay	6pm invoice saal mein do baar aik ghanta move
Cost and execution limits	Har beat ki max duration, turns, retries, spend aur termination behavior	Bhatakti run raat mein maheene ka budget kharch kare

Kit home 2 ki entrance fee hai aur travel karti hai: har baad ka home yeh chhe mangta hai, kabhi scripts ki jagah service settings. Part 5 ka chautha sawal speed limit lagaye to "minimum safety gate" yehi table hai.

Aam alfaaz mein

Pehli move sab se choti. Setup rakhein, sirf loop shuru karne wali cheez badlein. Laptop kholne ki jagah cloud schedule clock par shuru karta hai. Ab koi nahin dekh raha, is liye chhe chote safety rules jorein taake failure khamosh na ho.

Gehraai mein: schedule move hua, human gate nahin

Move mein subtle trap hai. Laptop par human gate ki sahulat yeh thi ke aap wahan the. Escalation us window mein aati thi jise dekh rahe the. Home 2 mein screen ke qareeb hon ya nahin, loop chalta hai. Unseen escalations jama aur delayed decision ban sakti hain, jo kabhi default decision hota hai. Is liye home 2 move woh upgrade mangti hai jo laptop ne nahin manga: escalation channel desk nahin, aap tak aaye, message, mention, ya aapke naam ka issue. Loop course ka "get loud" verb wahan point ho jahan aap rehte hain. Gate move nahin hua; doorbell ko hona para.

Khud jaanchein

Pehli cloud-scheduled beat green hai. Teammate kehta hai migration done. Aik run par evals ki seekh aur khamoshi par is part ki seekh se batayein: "done" se pehle kaun si do cheezein missing hain?

Jawab dekhein

Pehla: aik green beat aik run ka fact hai, home ka nahin. Migration rate se prove hoti hai, is liye bar pehli raat nahin, baseline ke muqable mein full operating cycle aur das beats hai. Doosra: loud failure test nahin hui. Jaan boojh kar failure plant aur alarm apne tak aate dekhe baghair khamoshi ambiguous hai: baseline bhi, ya runner shuru hi na hua. Done: aik hafta green, alarm tested, escalations wahan jahan aap waqai dekhte hain.

---

Part 3: Managed Runtime

5. Home 3: aap definition bhejte hain, service worker chalati hai

Home 2 ne clock move ki. Home 3 control plane ko move karta hai aur execution plane ko bhi, magar sirf agar aap chunein. Managed runtime simple aur radical contract wali service hai: aap agent describe karte hain, yani model, system prompt, tools, connectors aur guardrails; service use operate karti hai. Loop, session state, retries aur 3am crash recovery: control plane vendor computers par vendor engineers chalate hain, aur default mein sandbox bhi. Aap agent loop operate karna chhor dete hain. Gird ki har cheez ab bhi aapki: calling schedule, credentials, event consumption, escalations, spend limits, aur outcome ghalat ho to incident. Definition ke author aur business system ke owner aap hain.

Course likhte waqt concrete misaal Claude Managed Agents hai, jo April 2026 mein public beta mein aya. Details badlein tab bhi shape seekhein, kyunke shape idea hai. Aap teen cheezein banate hain:

• Agent: definition, yani model, prompt, tools, guardrails. Aapki rules file aur reviewer prompt us form mein jo service rakh sake.
• Environment: band jagah jahan agent actions execute hote hain. Harness course ki fences ab local config file nahin, service-side object. Yehi execution plane hai aur choice hai: default vendor infrastructure ka cloud sandbox, ya custody-required work ke liye aapki controlled infrastructure par self-hosted sandbox. Dono mein fences aap set karte hain. Vendor ki production guidance least-privilege networking aur explicit allowed-hosts list hai, bilkul harness course jaisi.
• Session: ongoing work, apne preserved state aur append-only event log ke saath. Loop course ki beat, lekin laptop ki zaroorat na hone se pause, resume aur kai din survive kar sakti hai.

Aapki application, ya is scale par script, API se in se baat aur events stream karti hai. Aham architectural detail: is home mein Concept 2 ke planes visible aur separable hote hain. Sochne wala model aur action ka sandbox alag pieces, service se jure. Pehle courses mein coupled version tha: agent aur tools aik machine ki process mein. Decoupled shape kisi aur ko acting part scale par operate karne deti hai, aur production architectures isi taraf ja rahi hain. Yeh idea Mode 2 mein aapki design choice ban kar lautega.

Do boundary facts seedhe, kyunke vendors dheere kehte hain. Managed control plane vendor-specific hai: yeh Claude, Anthropic ke operation mein. Sandbox aapki infrastructure par ho sakta hai, lekin loop, sessions aur model path nahin. Neeche OpenCode tab open-source side ka sach batata hai. Aur managed runtime hosting wala SDK nahin: Agent SDK aur Managed Agents alag products; aik ke liye code doosre par deploy nahin. Code carry nahin hota. Part 4 batata hai kya hota hai.

Aapki owned cheez se service-held cheez tak tarjuma: loop ki job description agent prompt banti hai. Permission walls aur deny rules environment config. Har scheduled beat session hai jo script headlessly usi schedule par kholti hai jo home 2 ne sikhaya. Endpoints, SDK call shapes aur beta headers book ki sab se mechanical layer hain. Chand hafton purane aur mahana badalte. Har detail live docs.claude.com se lein, aur season se purane har tutorial, isay bhi, shape samjhein, reference nahin.

Is side first-party home 3 nahin, aur course fake nahin karega. Magar missing cheez precise samjhein. OpenCode remotely chalta hai. opencode serve headless server process deta hai. Clients remote instance se attach aur sab managed compute par containerize ho sakta hai. Missing first-party managed control plane hai: vendor loop, sessions aur recovery service ki tarah operate kare. Interchangeable models wale open tool ka aik vendor nahin. Honest equivalents: hardened home 2, scheduler + kit aur vendor aap; ya aapka likha aur operate kiya server process. Serious use mein woh home 4 aur Mode 2 hai. Badle mein woh milta hai jo managed plane nahin bechta: aap aur loop ke darmiyan kuch nahin. Sahi trade tool ka nahin, Part 5 ka sawal hai.

Aam alfaaz mein

Homes 1 aur 2 mein worker employ bhi aap aur office maintain bhi aap. Home 3 mein job description aur office rules aap likhte, building-services company power, security, night shifts aur repairs wala office chalati hai. Aap front door nahin, reports se visit karte hain.

6. Kya milta, kya dete, aur cost kya hai

Managed contract ko dono taraf imaandari se tolein.

Kya milta hai. Woh operations jo aap nahin chahte: harness banane walon ki maintained sandboxing; crashes aur kai din ke jobs survive karta session state; current model ship karne wale vendor ka tuned context management aur prompt caching. Is se model-runtime compatibility ka manual work kam hota hai. Magar behavioral drift solve nahin, aur coordinated model+harness updates regression attribute karna mushkil kar sakte hain, kyunke do cheezein saath hilin. Scheduled baseline run ka kaam poora baqi. Infrastructure pager unka, yani raat ko engineer jagane wali alarm. 3am restart contract se aapka masla nahin. Business-outcome pager aapka: kaam waqt par unki machines par ghalat ho to escalation aapki.

Kya dete hain. Teen cheezein, sab se halki se sab se bhaari tak. Visibility: aap machine khud nahin, service ka event log parhte hain. Depth-3 trace grading, yaani sirf jawab nahin balkay agent ke uthaye qadam check karna, ab log mein dastiyab maloomat par depend karta hai. Custody: aap ke prompts, fixtures aur kaam ke inputs aisi infrastructure par execute hote hain jo aap control nahin karte. Kuch data, professions aur regulators ke liye yehi aik baat faisla kar deti hai; koi feature list jawab nahin badalti. Portability: definition vendor ki shapes mein likhi hoti hai, is liye chhorte waqt woh move nahin balkay rewrite hoti hai. Part 6 isi ka mazmoon hai.

Cost kya hai. Bill ki aik nayi qisam, jo aaj ke exact numbers se zyada aham hai. Homes 1 aur 2 mein tokens aur runner ki cost hoti hai. Home 3 mein runtime khud meter hota hai, bilkul bijli ke meter ki tarah. Is waqt active session work taqriban aath cents per hour hai, idle time free hai, aur tokens aur web search jaisi metered extras alag hain. Is shape ke do natayij hain. Jo session thori der soch kar der tak soti hai usay zinda rakhna lagbhag muft hota hai; isi se long-lived sessions affordable banti hain. Aur jo loop bhatakti hai, yaani evals course ki depth-3 disease, woh waqt ke saath paisa bhi zaya karti hai. Is liye eval suite sirf quality gate nahin, cost control bhi hai. Numbers badlein ge; live pricing page check karein. Asal sabaq bill ki shape hai.

Aam alfaaz mein

Managed runtime trade hai. Vendor unwanted operations sambhalta: crashes, restarts, sandbox health. Aap kuch visibility, data custody aur asaan portability dete hain. Naya bill active work-hours ka, idle free. Bhatakta loop ab waqt ke saath paise zaya karta hai.

Gehraai mein: vendor updates aik saath madad aur nuqsan

Evals drift story mein unchanged harness ke neeche model badla aur baseline ne tilt pakra. Home 3 shape badalta, drift nahin hatata. Ab model aur harness vendor schedule par saath badalte, engineers tune karte; aam tor par chote problems kam, kabhi relied behavior aapki file ke baghair badalta hai. Defense wahi: scheduled full-set, committed baseline, drop par loud alert. Managed runtime operations burden hatata, measurement burden nahin. Book mein koi cheez woh nahin hatati.

Khud jaanchein

Colleague kehta hai: "Managed sessions hourly metered, is liye apne schedule se zyada mehngi." Do misses batayein: meter ki, aur meter ne kya replace kiya.

Jawab dekhein

Meter active runtime ginta, idle free. Har beat chand minutes kaam aur baqi neend wali session runtime line par taqreeban kuch nahin. "Per hour" existing har hour nahin. Replace kiya: home 2 ki price sirf tokens nahin, tokens + runner + aapke hours thi, configure, update aur failure par jagna. Honest comparison operator samet total cost. Hobby loop mein home 2, team ko serve karte das loops mein pager ki bhi price. Is liye cost Part 5 ka chautha input hai, pehla nahin.

---

Part 4: Move Khud

7. Suitcase test: discipline travel karti hai, mechanics nahin

Homes ki har move ko course aik packing sawal banata hai: suitcase mein kya jata, aur pahunch kar kya dobara banta hai?

Trilogy ne jo sikhaya woh travel karta hai. Spec: job aur "done". Mehnat ke anchors wali rubric. Har case ki origin line aur baselines wala golden set. Ratchet log, caught failures ka record, har aik permanent test. Maker-checker split, category bars, human gate. Common baat: kuch software nahin. Yeh likhe hue decisions hain. Decision ko farq nahin kaunsa computer enforce karta hai.

Vendor ya machine ka naam wali cheez travel nahin karti. Flags, output formats, paths, session state, aik runtime API ka code. Concept 5 ka SDK-versus-managed boundary iska teekha form. Aur cost assumptions, jo homes mein size ke saath shape badalti hain.

Yahi aakhir lock-in ka asal jawab bhi hai, is liye saaf kahein: aap ka portable asset discipline layer hai, aur us ki portability koi maujood cheez nahin balkay woh cheez hai jise aap maintain karte hain. Jab bhi koi rule repo ke bajaye sirf vendor-side setting mein rahe, koi eval case sirf service ke andar ho, ya koi bar tay ho magar likhi na jaye, suitcase ka bojh bahar fixtures mein chala jata hai. Movable rehne ki aadat aik jumlay ki cost rakhti hai: truth repo mein hai aur har home usi se configure hota hai.

Aam alfaaz mein

Ghar badlein to apni cheezein pack aur deewar se lagi lights chhor dein. Spec, rubric, cases aur bars aapki, suitcase mein. Flags, paths aur API shapes old house se lagi, peeche. Movable log ghar mein rehte hue bhi valuables suitcase mein rakhte hain.

8. Trust transfer nahin, dobara kamaya jata hai

Figure ke right panel ka aakhri item apna concept chahta hai, kyunke movers sab se zyada isi ko skip karna chahte hain. 35/36 aik system ka measurement tha: yeh config, harness, model, machine aur reachable tools. Move ne kai saath badle. New home old system ka qareebi relative, magar old number us system par tha jo ab nahin. Inherit nahin ho sakta. Comparison target rehta hai: new home ko milne wala standard, free label nahin.

Arrival protocol evals course ko jaan boojh kar replay karta hai; move regression discipline ka sab se bara change hai:

1. New home mein pehle full golden set chalayein. Smoke nahin, full. Runner aur home headless hain. Isi liye Concept 3 bridge.
2. Misses ko count se pehle category mein parhein. Tone case down chota. New reach mein injection case down emergency; aur move se pehle rubric ya fence fix.
3. Bars hold, phir home-label baseline. Old baseline delete nahin, target. Location se release bars nahin badalte. New home existing bars pass kare. Old rate se meaningful gap investigate, accept nahin. Phir recorded date, model, rubric version aur runtime wala baseline. Explanation ke saath dono history mein. Environment baseline badalta, acceptance standard khamoshi se nahin.
4. Dependence se pehle probation. Fixed period: kam az kam full operating cycle aur das successful beats, riskier/rare mein zyada. Old home available, beats roz new baseline se checked. Initial operational evidence kahein, uptime proof nahin. New track record move complete karta hai, old reputation nahin.

Aam alfaaz mein

Aap ka old score old setup par measure hua tha. Move ne setup badal diya, is liye score saath nahin aata. New home mein usay dobara kamayein: day one par poora test set chalayein, dekhein kya fail hua aur har failure kis qisam ka tha, wahi pass-bar rakhein jo pehle tha, new home ke naam ke saath fresh score likhein aur trust karne se pehle trial period tak usay dekhein.

Harness course ki bad night jaisi warning. New home ki reach nai: cloud runner ke credentials laptop se alag, managed environment ke tools local config se alag. Injection aur blast-radius cases old reach ke khilaf. Probation se pehle hard cases se poochhein: kya guarded failure yahan alag dikhti hai? Aam tor par nahin; kabhi haan ho to sawal har future move mein jagah bana gaya.

Khud jaanchein

Home 2 ke baad full-set 33/36, old baseline 35/36: aik clean-fix case flake, re-run green; baqi do misses wahi case, old laptop ki absolute path wali fixture. Concept aur suitcase se sort karein.

Jawab dekhein

Flake noise, re-run policy se recorded, gating nahin. Repeated miss agent regression nahin, suitcase error: absolute path mechanics thi aur fixture mein travel kar gayi. Case ko relative-path repair karein, system ko nahin. Set badla to usi commit mein re-baseline. Bars nahin badalte, repair history mein visible taake genuine regression "suite fix" mein na chhupe. Honest reading: new home baseline par; move ne suite portability bug pakri, isi liye first run hoti hai.

---

Part 5: Home Chunna

9. Chaar sawal

Course ka sab kuch order mein chaar sawalon mein simat jata hai. Pehle teen home chunte. Chautha move ka waqt.

Q1: User kaun hai? Honest jawab aap ho to jaldi rukhein: home 2 taqreeban hamesha ceiling. Zyada readers, zyada waqt, aur maafi ki baat nahin. Jawab mein doosre log, team, client, customer, aate hi loop ko absence, vacation, login survive karna hai, sawal Q2.

Q2: Kya own karna lazmi hai? Yeh nahin ke aap kya own karna chahte hain, balkay kya waqai own karna zaroori hai. Do planes is sawal ko teen saaf jawabon mein baantti hain. Kuch own karna lazmi nahin: fully managed runtime munasib hai, aur vendor ko infrastructure pager sambhalne dena compromise nahin balkay paisay ka durust istemal hai. Sirf execution aur data plane, yaani kaam aur jis cheez ko woh touch karta hai, loop nahin: self-hosted sandbox wala home 3 isi ka jawab hai. Custody aap ke paas rehti hai aur operations aap rent karte hain. Control plane bhi: agar prompts, sessions, model path ya agent ki product surface bhi aap ko own karni hai to zaroorat home 3 se aage owned runtime, SDK path aur Mode 2 ki taraf jati hai.

Q3: Kya koi answer ka intezar karta hai? Scheduled/background work, triage, reports, pipelines, schedules aur managed sessions mein. Screen par human requirements badalta, owner lazmi nahin. Predictable startup, streaming, cancellation, concurrency wala serving runtime chahiye. Managed service self-operated se behtar de sakti; runtime own karna khud nahin deta, model thinking time koi runtime nahin hatata. Ownership Q2. Q3 serving ki alag shape batata, Mode 2 ki; filhal pehchan kaafi.

Q4: Buri raat ki cost? Harness budgeting aakhri baar. Low radius: kit pass, jaldi move, probation mein harden. High radius: first unattended shift se pehle kit aur full suite, injection categories all-of-them. Q4 destination nahin, speed limit.

Aam alfaaz mein

Chaar sawal isi tartib se poochein. Isay kaun use karta hai: sirf aap ya doosray log bhi? Aap ko kya own karna lazmi hai: kuch nahin, sirf kaam, ya poora loop? Kya koi shakhs screen par jawab ka intizar karta hai? Aur buri raat ki cost kya hai? Pehle teen sawal home chunte hain. Aakhri sirf yeh tay karta hai ke aap kitni tezi se move karein.

10. Aik poori move, aur homes jaan boojh kar mix

Ab poora course aik baar usi morning triage loop par chalta dekhein jo is section ke doosray course se saath chala aa raha hai.

Sawalat. Q1 user aap aur pichle maheene se report parhte do teammates. Woh "plus" trigger. Loop login survive kare. Q2 koi must nahin; repo GitHub par, queue custody-restricted nahin. Q3 koi screen par wait nahin; beat baithne se pehle. Verdict home 2, GitHub Actions, repo-attached aur eval gate wahi CI. Q4 bad labels aur aik wrong escalation: annoying, recoverable, low radius. Speed: kit pass, ab move, do-week probation, weekdays aur das beats floor.

Move. Monday: workflow file (schedule: trigger, headless invocation, repo config checkout, exit check, waqai parhe channel mein loud failure) plus teen cheap controls: concurrency lock, missed-run heartbeat, per-beat limits. Suitcase check mein laptop path fixture repair, same commit re-baseline. Pehli full-set all categories bar. runtime: actions baseline commit. Do haftay lid closed, beats schedule, khamoshi baseline, beech planted failure, kyunke kabhi na suni alarm afwah. Beat ten pass, probation end, laptop schedule delete. Do track records, naya matter.

Mix. Ab dekhein ke poora system kaisa lagta hai, kyunke yeh jaan boojh kar one-home nahin hai. Loop home 2 mein rehti hai. Eval gate evals course ke mutabiq usi CI mein rehta hai. Heavy one-off jobs, jaise quarterly cleanup ya bara refactor, ab bhi home 1 mein interactively chalti hain jahan aap unhein dekh sakein. Agar woh teammates client ban jayein to Q1 dobara uthta hai aur sirf serving path ke liye home 3 guftagu mein aata hai. Home permanent wafadari nahin, har loop ke liye chaar sawalon ka jawab hai, aur sehatmand system aam tor par do ya teen homes mein phaila hota hai. Concept 7 ka aik jumla is mix ko uljhan se bachata hai: repo truth rakhta hai aur har home usi se configure hota hai.

Aam alfaaz mein

Home woh jagah nahin jahan aap hamesha ke liye bas jayein. Chaar sawalon ki bunyaad par har loop ke liye alag choice hoti hai. Zyada tar real setups aik waqt mein do ya teen homes istemal karte hain: daily loop schedule par, aur heavy one-off job laptop par jahan aap usay dekh sakein. Aik rule mix ko mess banne se rokta hai: truth repo mein rehti hai aur har home usi se setup hota hai.

Khud jaanchein

Is book ki duniya ke aik doosray loop par chaaron sawal dobara chalayein: Ayesha ka invoicing loop ab paanch clients ko serve karta hai jo invoices seedha receive karte hain. Un mein aik bank hai jo shart rakhta hai ke client data us infrastructure par rahe jise Ayesha ki firm control karti hai. Har sawal ka jawab kahan land karta hai, aur ghair-araam-deh magar imaandar conclusion kya hai?

Jawab dekhein

Q1: users clients hain, is liye home 1 se aage aur dependence ke liye "jab runner chal jaye" se bhi aage jana hai. Q2 faisla-kun hai, aur do planes sawal ko precise banati hain: bank ka must execution aur data ke bare mein hai, yaani kaam kahan chalta aur kya touch karta hai. Firm ki infrastructure par self-hosted sandbox ke saath managed control plane shayad bilkul yahi zaroorat poori kare: client data firm ki custody mein aur loop vendor ke operations mein. Faisla Ayesha ka nahin, bank ka hai. Agar must control plane, prompts, sessions aur model path tak bhi jata hai to sirf owned runtime jawab hai. Q3: invoicing scheduled background work hai, latency pressure nahin. Q4: real clients ko ghalat invoices ka radius bara hai, is liye unattended move se pehle kit aur full suite chahiye. Imaandar conclusion yeh hai ke koi aik home fit nahin. Bank ka raasta ya home 3 with owned execution plane hai, ya zaroorat zyada ho to SDK path. Dono surat mein Ayesha operator ke akelay configure karne ki had par aa gayi hai; book mein is had ka naam Mode 2 hai. Isi lamhay ke liye book ka fork maujood hai.

---

Part 6: Imaandar Rehna

11. Lock-in rate hai, ownership drift karti hai

Pehle ke baad har home ke saath do slow failures, dono be-elan.

Lock-in event nahin, rate hai. Koi day one par portability sign away nahin karta. Woh ahista leak hoti hai: vendor-side setting mein badla rule jo repo mein mirror na hua, service console ke andar add hua eval case, ya dashboard mein dobara tay ki gayi bar jis ke peechay koi commit nahin. Har cheez suitcase se fixtures mein khamoshi se chali jati hai. Kisi bhi waqt lock-in ka seedha paimana yeh sawal hai: agar yeh home is quarter gayab ho jaye to move ki cost kya hogi? Bachao wahi jumla hai: repo truth rakhta hai. Isay evals course ke Part 6 mein Goodhart audit ki tarah, schedule aur hold-out mindset ke saath audit karein. Har quarter "sirf repo se fresh home configure karein" ki practice portability ka hold-out set hai. Agar fresh home complete nahin hota to leak us waqt mil gayi jab woh abhi sirf aik item chauri thi.

Kuch leak na ho tab bhi ownership drift karti hai. Zyada nazuk failure files mein nahin, aap mein hota hai. Mahinon khamoshi se chalne wala home zehan mein chhoti si promotion pata hai: system measured hai se system theek hai tak. Evals course ne iska mechanical roop bataya tha, jab drifted model ke saath judge ka 95 apna matlab badal deta hai. Runtime version insani hai: baselines ab bhi green, schedule ab bhi khamosh, aur aap dheere dheere per-category report parhna, calibration dobara chalana aur vendor ki nayi capability par Concept 8 ka new-reach sawal poochna chhor dete hain. Discipline mein "phir yeh khud ko maintain karta hai" naam ka koi qadam nahin. Scheduled run system ko dekhti hai; usay parhne ki calendar reminder aap ko dekhti hai.

Aam alfaaz mein

Do ahista masle chup kar barhte hain. Pehla: agar sab kuch repo mein na rakhein to har chhoti setting ke saath waqt guzarnay par home chhorna mushkil hota jata hai. Doosra: mahino khamoshi se chalne wala home aap ko usay check karna chhurwa deta hai. Dono ka aik hi hal hai: repo ko single source of truth rakhein aur reports waqai parhne ke liye calendar reminder lagayein.

12. Koi home kya fix nahin kar sakta, aur aage kahan

Section ko usi tarah khatam karein jaise is ka har course khatam hua: imaandar boundary par. Better home yeh badalta hai ke agent kab kaam karta, usay zinda kaun rakhta aur 3am par kya hota hai. Yeh is baat ko bilkul nahin badalta ke agent kitna acha kaam karta hai. Weak spec Anthropic cloud par bhi weak hai. Uncalibrated judge aath cents per hour par bhi uncalibrated hai. Missing eval case duniya ke har runner par missing hai. Runtime decision isi liye section ka aakhri course hai ke pehle course ke tor par woh bekaar hota: jo kuch move kiya ja raha tha usay move ke qabil banna tha. Agar is course mein move asaan laga to wajah yeh hai ke hard part trilogy thi. Move sirf suitcase hai.

Aam alfaaz mein

Better home yeh badalta hai ke agent kab chale, usay kaun zinda rakhe aur 3am par kaun theek kare. Woh agent ko zyada smart ya correct nahin banata. Weak plan cloud mein bhi weak rehta hai. Acha kaam us spec, harness aur tests se aata hai jo aap pehle bana chuke hain. Home sirf tay karta hai ke system ko chalta kaun rakhega.

Ab section ka poora arc saamne hai. Aap ne general agent ko drive karna, spec se direct karna, phir loop delegate, harness harden aur checker measure karna seekha, aur ab poore system ko munasib jagah house kiya. Aakhir mein aap ke paas wahi cheez hai jo book pehle page se khamoshi se bana rahi thi: specified, guarded, measured aur housed unit of work. Book ki vocabulary mein is jumlay ko rakhein to agla darwaza khud khulta hai: jab yeh unit doosray logon ke liye bani ho aur us ka owned runtime, product surface aur price ho to is ka naam Digital FTE hai. Aap poore safar mein operating scale par isi ko manufacture kar rahe thay.

Is section se teen darwazay nikalte hain aur chaar sawal pehle hi bata chuke hain ke aap kaunsa lein. Agla Personal Agent Harnesses section us reader ke liye hai jis ka Q2 personal scale par must own tha: worker, infrastructure aur poora runtime decision apne haath mein. Mode 1: Problem-Solving maujooda system se abhi real problems solve karne ke liye hai aur zyada readers ka durust agla qadam hai. Mode 2: Manufacturing mein is course ke tamam waday poore hote hain: Agent SDK se home 4 khulta hai, decoupled architecture virsay ke bajaye aap ki design choice banti hai aur Eval-Driven Development course us suite ko scale karta hai jis ne abhi aap ka move supervise kiya.

Aakhri khayal: loop ne agent ko waqt diya, harness ne limits aur evals ne track record. Is course ne worker ko aakhri cheez di: aik address jo aap ka nahin. Rhetoric ke neeche literal baat yeh hai ke independent address is section ki aakhri requirement hai, production operations ki nahin. Served product ko ab bhi owner, escalation path, retention policy, continuity aur usay band karne ka ikhtiyar rakhne wala shakhs chahiye; Mode 2 yeh sab sikhata hai. Magar pohanchna, kaam karna, kaam sabit karna aur apni jagah rehna hamesha poori job description thi. Agents ke liye, aur sirf agents ke liye nahin.

Khud jaanchein

Aik reader course khatam kar ke kehta hai: "To endgame home 3 hai. Aakhir sab kuch managed hi hota hai." Q1 se Q4, Concept 10 ke mix aur ooper wali boundary ko istemal kar ke do jumlon ki correction likhein.

Jawab dekhein

Koi endgame home nahin hai: chaar sawal har loop ke liye alag pooche jate hain aur sehatmand system jaan boojh kar kai homes mein phaila hota hai. Home 1 mein build, home 2 mein schedule, aur jab koi must talab kare to home 3 ya owned runtime se serve karein. Koi home agent ka upgrade bhi nahin: quality us spec, harness aur suite mein rehti hai jo saath travel karte hain. Home sirf tay karta hai ke lights on, yaani system ko chalta, kaun rakhega.

---

Is book ke apne loops kahan rehte hain (dogfooding)

Is course ka decision course ke wajood se pehle isi book par chal chuka tha. Book ka review loop, reviewer rubric, 95 bar aur is se neeche no-merge rule, Q1 ka jawab "aik team" deta hai: authors aur issues file karne wale readers. Is liye designed home home 2 hai, book ke repo par CI, aur bilkul wohi GitHub Actions version jo course sikhata hai, kyunke loop repo-attached hai aur eval gate ko merge gate ke paas rehna chahiye. Heavy interactive work, jaise naya course draft karna ya figure pipeline chalana, ab bhi home 1 mein hota hai jahan koi shakhs dekh raha ho. Is waqt home 3 aik khula sawal hai aur book usay bilkul Part 5 ki tarah treat karti hai: serving path ko koi must own majboor nahin karta aur latency pressure nahin, is liye managed option Q1 ka jawab dobara barhne tak intizar karta hai. Repo truth rakhta hai; homes details hain. Aap isi arrangement ka output abhi parh rahe hain.

---

🚀 Projects

Aasan se mushkil aath moves. Do rules: throwaway repo, failure khud plant. Home us bad night se proven jise survive kare.

📁Project 130-45 minHeadless wrapperAik beat command se chalayein, failure miss karna namumkin.

Difficulty: easy · Uses: Concept 3.

Build. Apne loop ki beat ko aik script mein wrap karein: headless invocation chalayein, exit code check karein, aur failure ko aisi jagah zor se report karein jahan aap waqai dekhte hain.

Done when network cable nikaal kar isay chalane par khamoshi ke bajaye alarm aaye. Is course mein aage kuch bhi mehfooz hone se pehle khamoshi ka matlab success hona chahiye.

⏰Project 21-2 hrsPehli scheduled beatSirf clock move karein, lid band beat dekhein.

Difficulty: easy to medium · Uses: Concept 4.

Build. Wrapper ko home 2 mein schedule par rakhein, ya to Routine se ya Actions ke schedule: trigger se, aur config repo se lein.

Done when aap ka laptop band hone ke bawajood aik beat chal chuki ho aur aap bata sakein ke result kahan gaya. Kit ki pehli teen rows bhi kisi na kisi shakal mein maujood hon, chahe chhoti hi kyun na hon: idempotency, missed-run detection aur concurrency lock.

🧳Project 345-60 minSuitcase auditDiscipline mein chhupi mechanics dhoondein.

Difficulty: medium · Uses: Concept 7.

Build. Do-panel wali figure khol kar apni config, cases aur fixtures ka jaiza lein. Discipline layer mein chhupi mechanics ki har cheez list karein: absolute paths, machine names aur prose mein jama diye gaye flags.

Done when list commit ho aur sab se buri teen cheezein theek ho chuki hon. Kam az kam aik milne ki tawaqqo rakhein; Part 4 ki kahani banayi hui nahin thi.

📏Project 41-2 hrsArrival protocolNew home full set aur imaandar re-baseline.

Difficulty: medium · Uses: Concept 8.

Build. Naye home mein poora golden set chalayein. Jo cases miss hon unhein category ke mutabiq sort karein. Runtime ke naam ke saath nayi baseline record karein.

Done when baseline.json mein runtime: field ho aur har miss ka likha hua verdict maujood ho: noise, suite bug ya real. Real failures ka fix ship ho chuka ho.

🌙Project 51 hr, plus two weeks of nightsProbationDas beats, planted failure, old home delete.

Difficulty: medium · Uses: Concepts 4 and 8.

Build. Naye home mein kam az kam aik poora operating cycle aur das scheduled beats chalayein, har roz nayi baseline ke muqablay mein check karein aur old home ko dastiyab rakhein. Darmiyan mein jaan boojh kar aik failure lagayein.

Done when das beats green hon, jaan boojh kar lagaya gaya alarm aap tak pohanchay aur old schedule delete ho chuka ho. Course ke mutabiq isi ka matlab "moved in" hai. Log ke ooper "initial operational evidence" likhein, kyunke das beats sirf itna hi sabit karti hain.

🗺️Project 645-60 minChaar sawal, likh karHar loop ka Q1-Q4 jawab commit karein.

Difficulty: medium · Uses: Concepts 9 and 10.

Build. Har us loop ke liye jo aap waqai chalate hain, aik committed Markdown file mein Q1 se Q4 tak jawab dein aur har jawab ke aakhir mein home aur speed limit likhein.

Done when sirf woh file parhne wala shakhs bata sake ke har loop kahan rehta hai aur kyun. Kam az kam aik loop ka jawab itna hairan-kun ho ke aap usay move karein.

🏢Project 72-3 hrsHome 3 mein aik sessionManaged session kholein, event log ki had dekhein.

Difficulty: medium to hard · Uses: Concepts 5 and 6. (Claude Code track. OpenCode readers: Project 2 runner harden: credentials rotated, updates scheduled, uptime checked.)

Build. Live docs ki madad se aik managed agent, aik environment aur aik session banayein jo aap ke golden set ka aik graded case chalaye. Event log ko shuru se aakhir tak parhein. Active-runtime meter ko session se pehle aur baad mein note karein.

Done when aap apne run ki bunyaad par bata sakein ke log ne kya dikhaya, kya nahin dikha sakta tha aur session ki cost kya thi. Yeh teen jumle Project 6 ki file ke saath commit hon.

🔥Project 82 hrs, then a quarter of patienceVanishing-home drillRepo alone se home dobara, time hi lock-in.

Difficulty: capstone · Uses: Concept 11.

Build. Sirf repo se, vendor console ya old machine ke baghair, apne loop ke home ki fresh copy configure karein aur usay baseline tak le jayein. Waqt note karein aur drill ko har quarter schedule karein.

Done when fresh home poori suite pass kare aur likha hua time-cost aap ka measured lock-in ho. Agar drill fail ho to aap ne leak us waqt pakar li jab woh sirf aik item chauri thi; poora maqsad yahi hai.

---

Sources aur mazeed reading

Is book mein

• Loop Engineering: beats, schedule, "get loud".
• Harness Engineering: fences aur Q4 blast-radius budget.
• Checker par Bharosa: golden set, baselines, drift gate.
• Personal Agent Harnesses: Q2 ka personal must own.
• Mode 2 mein Eval-Driven Development aur Deploy the Agent Harness: manufacturing size, home 4 open.

Runtimes (official docs)

• Claude Code headless: claude -p, formats, scripting: https://code.claude.com/docs/en/headless
• Claude Agent SDK overview, SDK-versus-managed: https://code.claude.com/docs/en/agent-sdk/overview
• Claude Managed Agents quickstart: agents, environments, sessions: https://platform.claude.com/docs/en/managed-agents/quickstart (April 2026 public beta, fastest-moving docs)
• Managed environments, cloud sandboxes, least privilege, self-hosted: https://platform.claude.com/docs/en/managed-agents/environments
• Managed pricing, active session-hour meter: https://platform.claude.com/docs/en/about-claude/pricing
• Claude Code Routines, /schedule, triggers, stagger, green caveat: https://code.claude.com/docs/en/routines (research preview)
• Cowork Scheduled Tasks aur local caveat: https://support.claude.com/en/articles/13854387-schedule-recurring-tasks-in-claude-cowork
• OpenCode CLI/server: opencode run, opencode serve, remote attach: https://opencode.ai/docs/cli/
• GitHub Actions schedule: aur delay/drop: https://docs.github.com/en/actions/using-workflows/events-that-trigger-workflows#schedule aur https://docs.github.com/en/actions/how-tos/troubleshoot-workflows

Sab links mid-July 2026 current. Mechanical layer sab se tez purani. Rely se pehle live docs par name, endpoint, price confirm.

---

Aik-line summary

Aik sawal se shuru karein: loop kaun operate karta hai aur kaam kahan execute hota hai? Headless mode har naye home tak bridge hai. Discipline travel karti hai, mechanics dobara banti hain aur move ke baad trust phir measure hota hai. Chaar sawal home chunte hain aur blast radius move ki raftaar tay karta hai. Homes har loop ke liye jaan boojh kar mix hote hain, aur repo truth rakhta hai taake aap hamesha nikal sakein. Worker tab mukammal hota hai jab us ka address aap ka na ho.

Flashcards Study Aid

---

Apni Samajh Jaanchein

Checking access...