Workflows & model ka intekhab
AI agent ke ird-gird kaam ko organize karne ka tareeqa bohat tezi se badla hai — har diff ko micromanage karne se le kar raat bhar Ralph Loop chalane tak. Yeh hain woh 6 workflows, sahi tareeqa kaise chunein, aur models ka imandaranah mawazna kaise karein.
01Workflows ka irteqa (evolution)
Aik workflow yeh hota hai ke aap agent ke ird-gird apni sargarmion ko kaise organize karte hain. Karpathy ka tweet isi bare mein tha, aur yeh dono mindsets ke sath tabdeel hua hai. Trust ke 6 levels ke tor par sochein — jo kal ke 8 stages se milti julti hain.
2025 mindset — trust ke 3 levels
1 · Micromanagement
agents.md mein bohat specific instructions, har change ko approve karna, bar bar rokna, rewrite karna, reset karna aur dobara chalana. Mohtat aur hands-on tareeqa — jahan se aksar logon (aur instructor) ne shuru kiya tha.
2 · Plan → execute → review → test
Plan mode use karein: agent to-do plan likhta hai; aap razi hote hain, phir execution par switch karte hain. Phases mein build karein — har phase ke liye: build, code-review (usay chhoti baat ka batangarr banane se rokein), test, results review, complete mark karein, aur aage barhein. Thora zyada ikhtiyar, magar control ab bhi aapke paas.
3 · Spec-driven development (“trust but verify”)
Wazeh taur par specify karein ke kya chahiye — baaz auqat formal spec language mein — phir usay chalne dein. Aap har qadam ko dekhne ke bajaye aakhir mein verify karte hain. Nigrani mein aik wazeh kami.
2026 mindset — mazeed 3 levels
4 · YOLO
Koi approvals nahi — agent jo chahe kare, bina permissions ke. “Trust but verify” ko agle level par le jana: start karein, khana khane chale jayein, wapis aakar dekhein usne kya banaya. 2025 ke aksar hissay mein yeh sirf shauqiya tha; ab log real kaam ke liye use karte hain.
5 · Ralph Loops
Ralph Wiggum (masoom, pur-umeed) ke naam par aur Geoffrey Huntley ki ijad. Agent pehle hi andar loop karta hai; Ralph Loop is poore loop ko aik bare loop mein lapet deta hai: run karein, test karein “kya yeh kaafi hai?”, feedback generate karein, usay objectives mein add karein, aur dobara chalayein — taqreeban 10 bare loops tak. Raat ko chalayein; subah aakar dekhein bohat sara kaam mukammal ho chuka hoga.
6 · Multi-agent / swarms
Mukhtalif roles wale muta’addid agents — testing, feedback, hierarchy mein manager aur worker agents. Badi tadaad mein spawn aur orchestrate kiye gaye. Yeh 2026 ki frontier hai.
02Kaunsa tareeqa sahi hai?
Classic jawab: koi aik tareeqa “sahi” nahi hai — mukhtalif tasks ke liye mukhtalif tareeqay munasib hain. Isay do hisson mein baantein:
Hands-on rahein (levels 1–3)
Enterprise software, commercial SaaS, bade codebases, ya intehai innovative code (maslan naye MCP servers jinhein models mushkil se handle karte hain). Precision aur nigrani sab se ahem hain.
Chalne dein (levels 4–6)
MVPs, prototypes, pilots, bilkul nayi empty directories, bohat sa boilerplate (HTML, React app, CRUD backends), aur thora risk lene ka irada. YOLO aur Ralph Loops ke liye behtareen.
Amali taur par instructor zyada tar pehli category mein rehte hain — mission-critical, aksar naya kaam jaise MCP integrations, jahan mojooda LLMs non-idiomatic code likhte hain kyunki pattern abhi naya hai. Magar kal ka game doosri category mein tha: shuru mein bunyadi taur par YOLO, phir fancy version ke liye Ralph Loop. Yeh course mission-critical, bade codebase aur real enterprise patterns par focus karta hai — sath hi 2026 techniques ka bharpoor jaiza leta hai.
Aapka kaam aisa code deliver karna hai jo sabit shuda kaam karta ho. “LLM ne likha tha” koi bahana nahi hai. LLMs ko khul kar use karein — yeh aapko bohat zyada kaam karne dete hain — magar task ke mutabiq sahi tareeqa chunna, check karna aur validate karna ab bhi aapki zimmedari hai. Jo code aap ship karein uski accountability lein, chahe AI ne aapki madad ki ho ya na ki ho.
03Hype se aage ki haqeeqat
Aik sachay shauqeen ki taraf se aik sanjeeda aur imandaranah baat. AI ki dunya mein bila-waja ki hype mojood hai.
Hisaab se kaafi zyada tez
Wazeh target ke sath boilerplate-heavy kaam — bohat se components wala React front end. Aik achay front-end dev ke liye aik din ke bajaye minute; greenfield projects ke liye dinon ka kaam minute mein.
Sirf mamooli tez — ya ulta sust
Bade codebases par intehai naya aur complex kaam. Baaz auqat isne kaam ko sust bhi kiya hai — aik aisi bareek aur ghair-mutawaqqo ghalti jise dhoondna aur dobara likhna pada.
Khulasa: LLMs waqai aik taqatwar multiplier hain — magar har jagah flat 10× nahi. Faida kitna hoga iska daromadar project par hai. Hamara kaam tools ko seekhna, unhein use karna aur haqeeqat wazeh karna hai — un logon ke liye jo over-hype hain aur unke liye bhi jo anti-AI hain — yeh batana ke AI kahan behtareen hai aur kahan isay abhi behtar hone ki zaroorat hai.
04Models ka mawazna — artificialanalysis.ai
Agar aapko sirf aik bookmark rakhna ho, to artificialanalysis.ai ko rakhein. Yeh models ka intelligence, speed aur price ke lehaz se mawazna karta hai, har us tareeqe se jo aap dekhna chahein.
Top models late-2025 ke inflection point ke ird-gird jama hain jab coding agents ki reliability mein aik bada jump aaya. Kuch skeptics ke zehan mein abhi bhi pre-November daur ke bure tajrubat hain — unhein dobara try karne ka mashwara dein kyunki ab bohat kuch badal chuka hai. Isi site par aapko frontier-intelligence-over-time chart bhi milega (yeh tezi se oopar jati curve ki tarah dikhta hai; instructor ka khayal hai ke yeh haliya jump reasoning techniques ki wajah se aaya aur hamesha itna sharp nahi rahega — check karein ke yeh peshangoi sahi sabit hoti hai ya nahi), sath hi coding- aur tool-use ke makhsoos tests, jahan open-source Z.ai ka GLM-4.7 hairan kun taur par mazboot score karta hai.
Aapne is poore course mein yahan sab se zyada baatein suni hain — shukriya. Ab aap samajh chuke hain ke engine ke andar kya hai. Kal: tools, tools, tools — Cursor, Copilot, Codex, Antigravity — aur aapka pehla real build. Aap safar ka taqreeban 13% hissa tay kar chuke hain. 🎉
✓ Ahem takeaways
- 6 workflows micromanagement se swarms tak phailay hue hain, 2025 aur 2026 ke mindsets ke sath.
- Ralph Loops agent ke loop ko aik bairooni feedback loop mein lapet dete hain — ghanton ya raat bhar chalayein.
- Workflow ko task ke mutabiq match karein: mission-critical ke liye hands-on, greenfield/boilerplate ke liye chalne dein.
- LLMs aik real multiplier hain, har jagah flat 10× nahi — aur “LLM ne likha tha” kabhi bahana nahi hota.
- Intelligence, speed aur price ka mawazna karne ke liye artificialanalysis.ai ko bookmark karein.