General Agents

Woh tools jo sirf baat nahin karte: kaam karte hain. In mein se aik ko drive karna seekhein; book mein aage ki har cheez isi par banti hai.

Foundations ne aap ko sikhaya ke AI kya hai aur use istemaal karte hue apni judgment kaise bachani hai. Yeh section woh jagah hai jahan aap woh tool uthate hain jo aap baqi book mein istemaal karenge: aik general agent, yani aisa AI jo sirf answer nahin karta, balkay act karta hai. Yeh aap ki files kholta hai, unhein parhta hai, code likhta aur run karta hai, aur task complete karne ke liye doosri apps bhi istemaal karta hai.

Shuru karne ke liye kuch install karna zaroori nahin. Is section ka pehla qadam usi browser tab mein hota hai jahan Foundations chali thi:

• Web par General Agents - poore section ka front door. Browser tab se apna pehla real agent direct karein, kuch install kiye baghair: remote session kya hai, aapki files asal mein kahan rehti hain, approval aapke phone tak kaise pahunchta hai, aur har naye agent product ko ek shared shape se kaise parhna hai. Har koi pehle din yahan se shuru karta hai, chahe code kar sakta ho ya nahin.

Jab agent ko apni machine par lane ke liye tayyar hon, baqi section ke do darwaze hain, aur kaunsa lena hai yeh sirf is par depend karta hai ke aap kaun hain:

• Claude Code aur OpenCode - un logon ke liye jo code ke saath kaam karte hain.
• Cowork aur OpenWork - baqi sab ke liye: wahi power desktop app mein, jo professional aur knowledge work ke liye bani hai.

Idea wahi hai, audience alag hai: aisa AI jise aap real work karwane ke liye direct karte hain, na ke aisa AI jis se aap sirf chat karte hain.

Yeh poori book ka hinge hai. Is ke baad aane wali lagbhag har cheez yeh maan kar chalti hai ke aap general agent drive kar sakte hain. Mode 1 (aik problem aik dafa solve karna) aur Mode 2 (permanent worker manufacture karna), dono asal mein "aap aur aik general agent" hain. Optional Personal Agent Harnesses bridge bhi apna harness inhi tools mein se kisi aik ke through install aur run karta hai. Is liye yeh woh skill hai jise aap sab se zyada reuse karenge. Baad mein, yehi general agent jise aap yahan drive karte hain, woh tool ban jaata hai jise aap workers build karne ke liye istemaal karte hain. Tool wahi, kaam bara.

Yeh section teen stages mein

Yahan ke courses aapko aik line par aage le jate hain: tool ko har second pakar kar chalane se, use wazeh tor par direct karne, aapke liye chalne wala loop design karne, us loop ke harness ko harden karne, kai loops ko shared memory dene, aur phir naapne tak ke center ka checker bharose ke qaabil hai ya nahin.

Stage 1: Drive

General agent ko achhi tarah operate karna seekhein. Har koi web se shuru karta hai, kuch install kiye baghair; phir apne liye sahi installed-tool course karein aur baqi baad mein utha lein.

• Web par General Agents - yahan se browser mein shuru karein, kuch install kiye baghair. July 2026 ke web work surfaces, Claude Cowork on the web aur ChatGPT Work, par apna pehla agent direct karein: remote sessions, files kahan rehti hain, phone tak pahunchne wali approval, aur har agent product ko samajhne ke liye six-part lens. Har koi pehle yeh karta hai, code kar sake ya nahin; phir neeche se installed tool chunta hai.
• Open Source LLMs: Aapka Laptop, Aapka Server/Cluster, Cloud - coding track ka pehla stop. Open models free download hote hain, is liye asal sawal yeh hai ke unhein kahan chalaya jaye. Yeh course teen scales par jawab deta hai: Ollama ke saath apne laptop par, vLLM ke saath apne control wali GPU machine par jo aik waqt mein pachaas log serve kare, aur OpenRouter ke zariye cloud mein jahan woh frontier open models rehte hain jinhein koi shakhs self-host nahin kar sakta. Aik idea sab ko jorta hai: aapka agent harness plus swappable brain hai, aur brain bas aik address hai. Part 1 akela kaafi hai aur free install ke siwa kuch nahin chahta. Coding agent pehle drive kiya ho to madad milti hai, magar nahin kiya to iska one-command install aapke liye khara kar deta hai. Knowledge workers apni private AI ke liye Part 1 kar ke seedha Cowork ja sakte hain.
• Agentic Coding: Claude Code aur OpenCode - un logon ke liye jo code ke saath kaam karte hain. Plan mode, context management, rules file, skills, subagents, aur connectors (MCP): coding agent ko kaise drive karein taake woh aap ki files parhe, plan propose kare, edits kare, aur aap ko result check karne de.
• Professionals ke liye Cowork aur OpenWork - baqi sab ke liye. Wahi qisam ka agent desktop app mein, jo professional work ke liye bani hai (documents, spreadsheets, slides, research), taake usay kaam par lagane ke liye aap ko terminal chhoona na pare.
• Website Design - woh course jahan aap machine ke baare mein seekhna band kar ke use kisi asal kaam par lagate hain. Aik client ke liye aik website, khaali folder se live address tak jise koi aur phone par khol sake: pixels se pehle plan ki hui kahani, client brand ko mood ke bajaye rule samajhna, asal photos aur video, phone version ko aakhri khayal na banana, aur screenshots ko tests ke tor par istemaal karna taake "kya yeh sahi lagta hai" waqai check ho sake. Iska aik idea woh hai jise Stage 2 aur 3 baar baar istemaal karte hain: taste ab aik file hai. Made page aur generated page ke darmiyan judgment likhi ja sakti hai aur koi bhi model use load kar sakta hai. Oopar ke kisi tool ko drive kar sakne ke baad ise karein; kuch ship karne aur kaam ko ghalat hota dekhne ke baad specs aur loops ki qadar zyada samajh aati hai.

Stage 2: Direct

Drive karne se result milta hai; direct karne se har dafa sahi result milta hai.

• Spec-Driven Development - agent ko vague instructions dena band karein aur usay likhi hui spec dein: exactly kya karna hai, kis par karna hai, aur "done" ka matlab kya hai. Clear spec woh farq hai jo guessing agent aur target hit karne wale agent ke darmiyan hota hai, aur Mode 2 mein cross karte waqt yehi pehli cheez aap saath le jate hain.

Stage 3: Loop delegate karein

Aap is stage ki shuruaat aik naqshay se karte hain, phir us naqshay mein naam di hui cheezein banate hain: loop design karein, jis harness mein woh chalta hai use mazboot karein, kai loops ko aik mushtarka memory dein, phir evals se sabit karein ke darmiyan ka checker bharosemand hai. Aakhri course us sawal ka jawab deta hai jo yeh sab uthate hain: jab loop par bharosa ho jaye to use kahan rehna chahiye?

• The Four Layers - ise pehle parhein, aur aik ghante se kam mein parhein. Yeh aage ki har cheez ka naqsha hai: prompt, context, harness aur loop chaar mukhtalif skills nahin jin mein se aap aik chunein, balke chaar containers hain, har aik agle ke andar aur har aik ki apni unit of work hai. Is ka asal faida technique nahin, aik aadat hai. Jab agent chalees minute aur pachaas dollar aik hi kaam dohrane mein zaya kare, to aap prompt dobara likhna band kar dete hain, jise das seconds mein edit kiya ja sakta hai, aur poochte hain ke asal mein kaunsi layer tooti. Yeh yeh bhi wazeh karta hai ke graphs kahan fit hote hain: paanchveen layer ke taur par nahin. Kuch install karne ki zaroorat nahin; aik chat tab aur aik agent jise aap pehle chala chuke hain, bas itna hi setup chahiye.
• Loop Engineering - tool pakar kar rakhne se aisa system design karne tak ki chhalang jo agent ko aap ke liye prompt karta hai. Aap aik chhota loop banate hain jo wake up karta hai, dekhta hai kya badla, tay karta hai kya karne layak hai, har job agent ko deta hai, result check karta hai, aur sirf un decisions ke liye aap ko bulata hai jinhein waqai insaan ki zaroorat hai. Valuable skill aap ke likhe prompt se hat kar aap ke design kiye loop mein aa jaati hai. Yehi aage ki har cheez ka natural doorway hai.
• Harness Engineering - usi chhalang ka bharose wala hissa. Aisa loop jo aap ke sote hue chalta rahe, usay aik aisi layer chahiye jo tay kare ke agent kya kar sakta hai, usay kya jaanna chahiye, us ka kaam kaise saabit hota hai, aur ghalat hone par kya hota hai. Wohi layer harness hai, Claude Code aur OpenCode dono ke paas apna apna harness hai, aur yeh course aap ko sikhata hai ke usay soch samajh kar kaise engineer karna hai: permission walls, sandboxes, automatic checks, aur woh habit jo har pakri gayi ghalti ko aise rule mein badal deti hai jo usay dobara hone se namumkin bana deta hai.
• Graph Engineering - memory wala hissa, aur woh course jo aik loop kaafi na rehne par karein. Bees agents fan out karein to har aik khaali shuru hota hai, jo doosre ne aik ghanta pehle dhoonda use dobara dhoondta hai: kaam multiply hua, memory nahin. Hal yeh hai ke findings transcripts mein chhorna band karein aur unhein typed, connected records ke tor par likhein jinhein baad ka koi agent query kar sake, yani agent bhoolta hai, graph nahin. Do graphs rakhein aur kabhi merge na karein, aik tried work ka aur aik established facts ka; har claim ke saath receipt lagayein; har worker ko dump ke bajaye chhota bounded slice dein; aur reviewer ko opinion ke bajaye edge cite karne par majboor karein. Aakhri part loops ko khud wire karta hai: kaun kis ko check karta hai, kis ke target ka owner kaun hai, aur kaun se measurements par koi loop behes nahin kar sakta. Doosra loop banate waqt ise parhein, aur yeh imaandaari se batata hai ke graph kab nahin banana.
• Checker par Bharosa - loop chalta hai, harness kaam prove karta hai, aur center mein checker PASS ya FAIL kehta hai. Lekin kaise pata ke checker acha hai? Yeh course evals sikhata hai: real failures se test cases ka chhota folder banayein, reviewer ko aik baar ke bajaye kai baar grade karein, apni judgment ke saath calibrate karein, aur har change ko result par gate karein. Sirf files, shell aur jq se "checker ne PASS kaha" ko aise number mein badalta hai jiska aap difa kar sakein.
• Laptop Chhorna - runtime decision. Laptop par proven loop aik machine mein phansa hai; yeh course us track record ko khoye baghair use wahan le jata hai jahan woh asal mein chale, cloud schedule, managed runtime ya aapka apna process. Yeh Stage 3 close karta hai: loop chalta hai, harness guard karta hai, checker trusted hai, aur ab woh aisi jagah rehta hai jo aapke sone par nahin soti.

Chaar Alfaaz, Aik Doosre Ke Andar

Oopar ki teen stages chaar aise alfaaz istemaal karti hain jinhein industry baar baar mila deti hai: prompt, context, harness, aur loop. Yeh un ki shakal hai, taake aap dekh sakein ke kaunsa course kis ka zimmedar hai. The Four Layers woh course hai jo taqreeban pachaas minute mein ise theek tarah sikhata hai, aur Stage 3 mein sab se pehle isi ko parhna chahiye.

Lafz	Yeh kya hai	Kaam ki unit	Iska owner course
Prompt	Aapka bheja message: ask, examples aur format.	Aik message	2026 mein AI Prompting
Context	Aik response ke liye model jo kuch dekhta hai: files, history, rules, tool results.	Window mein qadam-ba-qadam rehne wali cheez	Agentic Coding, Parts 2-4
Harness	Aik run ke andar rules: agent kya kar sakta hai, kaam kaise prove hota hai, fail ho to kya hota hai.	Aik run	Harness Engineering
Loop	Har cheez ke gird schedule: runs start aur grade karta hai, unke darmiyan yaad rakhta hai.	Kai runs, kai din	Loop Engineering

Yeh chaar line mein steps nahin. Yeh aik doosre ke andar layers hain. Acha prompt bad context mein fail hota hai. Acha context bare harness mein fail hota hai. Acha harness use chalane wale loop ke baghair idle rehta hai. Isi liye demo mein chalne wala agent production mein aksar mar jata hai: demo ko sirf andar ki layers chahiye thin aur bahar ki layers kabhi banin hi nahin. Model masla nahin tha. Uske gird layers ghaayab thin.

Is section ka aik course in chaaron mein se kisi ka owner nahin, aur yeh jaan boojh kar hai. Graph Engineering paanchveen layer nahin, balke aik topology hai: aise kai stacks jo aik doosre se jure aur aik memory share karte hain. Woh table mein is liye nahin ke jab tak aap aik se zyada loop na chala rahe hon, yeh lafz koi faida nahin deta. Isi liye woh chaaron ke baad aata hai.

Acha prompt aapko hoshiyar jawab deta hai. Chaaron layers mil kar aapko woh cheez deti hain jo aap ship kar sakein. The Four Layers har layer ko alag karta hai, har aik ki failure signature deta hai, aur "mera agent kharab hai" ko aise sawal mein badalta hai jis ka aik pata ho.

Pehle aap ko kya chahiye

Pehle Foundations section karein, khaas taur par AI ke Daur Mein Sochna (apni judgment bachana) aur Skills aur Connectors (woh plugins jin se general agent aap ke tools aur data tak pahunchta hai). Aap ko programmer hona zaroori nahin: agar aap nahin hain, to Cowork aur OpenWork wala darwaza lein aur aap theek rahenge.

Yeh kahan le jaata hai

Jab aap general agent ko drive, direct, loop, aur harden kar sakte hain, to baqi book khul jaati hai. Optional Personal Agent Harnesses section bas aage hai, agar aap aisa worker own karna chahte hain jo aap ki apni infrastructure par chalta hai. Phir woh fork aata hai jis par poori book ghoomti hai: Mode 1 - Problem-Solving, jahan aap aik problem aik dafa solve karne ke liye general agent istemaal karte hain, aur Mode 2 - Manufacturing, jahan aap permanent worker build karte hain jo usay hamesha solve karta hai.

Book ka arc aik line mein: Foundations se samajhna, General Agents se drive karna, aur do modes se kaam mein lagana.

Browser mein Web par General Agents se shuru karein, har koi pehle din kuch install kiye baghair kar sakta hai. Phir jab agent ko apni machine par laayein to apna darwaza chunein: code ke saath kaam karte hain to Claude Code aur OpenCode, warna Cowork aur OpenWork.