Chaar Layers: Prompt, Context, Harness, Loop
12 concepts · taqreeban 50 minutes · kuch install nahin karna · woh chota course jis par aap baar baar lautenge
Aik agent chalis minutes chalta hai. Pachas dollars kharch karta hai. Aisi koi cheez nahin banata jo aap use kar sakein. Baad mein log parhein to dikhta hai: us ne wohi teen cheezein baar baar azmayin, har baar sirf chand lafz badle.
Aap kya badlenge?
Taqreeban har shakhs prompt badalta hai. Wohi pehli cheez hai jo nazar aati hai, aur das seconds mein edit ho sakti hai. Log usay dobara likhte, run phir shuru karte, aur agle chalis minutes wohi hota dekhte hain.
Prompt theek tha. System mein progress rukne ko notice karne wali koi cheez thi hi nahin, is liye kisi ne run ko roka nahin. Fix taqreeban gyarah lafz tha, doosri file mein, aisi layer par jis ka naam zyada tar log nahin jante.
Yeh course aapko naam deta hai: prompt, context, harness, loop. Yeh chaar skills ya chaar seedhiyan nahin. Yeh chaar containers hain, har aik agle ke andar, aur har aik ka kaam alag hai. Inhein dekhna aa jaye to "mera agent toot gaya" address wale sawal mein badalta hai: kaun si layer tooti?
Yeh jaan boojh kar is section ka sab se chota course hai. Yeh naqsha hai, aur jo naqsha yaad na rahe woh apna kaam nahin kar raha.
Aap ne kam az kam aik baar general agent chalaya ho, kisi bhi darwaze se: code ke saath kaam karte hain to Claude Code aur OpenCode, warna Cowork aur OpenWork. Spec-Driven Development madad karta hai magar lazmi nahin. Yahan repository, database, ya install ki zaroorat nahin. Aik chat tab aur pehle use kiya hua aik agent poora setup hai.
📚 Teaching Aid
Poori Presentation Dekhein: Chaar Layers: Prompt, Context, Harness, Loop
Yeh alfaaz abhi uljhan kyun paida karte hain
Online parhte hue yeh alfaaz phisalte lagein to masla aap nahin. Field nai hai aur isay vocabulary ghalat order mein mili.
Kuch arsa pehle sirf aik layer thi jise koi chhoo sakta tha. Aap message type karte aur jawab parhte. "Prompt engineering" poori skill ka naam tha kyun ke prompt hi poori surface thi.
Phir tools tezi se bahar ki taraf barhe. Coding agent ne permissions file, rules file aur hooks diye. Yeh model ke gird configurable layer hai. Phir schedules aur routines aaye. Yeh pehli ke gird aik aur configurable layer hai. Kam waqt mein do nai surfaces aa gayin.
Vocabulary saath nahin chali. Har surface ka naam pehle likhne wale ne rakha aur naam takra gaye. "Context engineering" kabhi window aur kabhi model run se pehle ka sab kaam hai. "Graph" teen cheezon ke liye use hota hai, jinhein Concept 11 alag karta hai. Sab se zyada mushkil harness deta hai. Online writing aksar harness aur loop ko aik layer kehti hai. Kuch writing tools, credentials aur security boundary dene wale platform ko harness kehti hai.
Yeh alfaaz ki be-zarar behas hoti, agar aik baat na hoti. Missing permission rule aur missing schedule alag bugs hain. Woh alag surfaces par rehte aur alag log fix karte hain. Donon ke liye aik lafz aapko ghalat jagah bhejta hai, bilkul opening story ki tarah.
Is liye course do kaam karta hai: chaar layers ko containers ke taur par deta hai, aur aisa test deta hai jo alfaaz dobara badalne ke baad bhi chalega.
Aasaan alfaaz mein key terms
Abhi aik baar parhein. Term dhundla ho to laut aayein. Har term baad mein poora samjhaya gaya hai.
| Term | Aasaan ma'ni |
|---|---|
| Prompt | Aap ka bheja message: ask, examples, format aur role. |
| Context window | Aik response likhte waqt model jo kuch dekh sakta hai. Jo is mein nahin woh fact ke taur par available nahin. |
| Curator | Jo tay karta hai window mein kya jaye, kis order mein, aur kya bahar rahe. |
| Beat | Agent ka aik poora turn: aap ki instruction se tamam tool calls aur phir quiet hone tak. |
| Harness | Model ke gird code jo aik beat chalata hai. |
| Loop | Harness ke gird system jo beats shuru, judge, aur un ke darmiyan yaad rakhta hai. |
| Heartbeat | Beat shuru karne wali cheez: schedule, event, ya condition. |
| Spine | Model se bahar saved state, taake agla beat pichle ka kaam jane. |
| Stopping condition | Testable rule jo kaam complete kehti hai. Maker ke ilawa koi ise chunta aur enforce karta hai. |
| Maker-checker | Aik agent kaam karta hai, doosra agent ya command check karta hai. |
| Human gate | Woh point jahan run aage barhne se pehle insaan faisla karta hai. |
| Sub-agent | Apni window wala helper jo aik job karke summary lautata hai. |
| Graph | Aise kai stacks ki topology: aage kya chale, har edge par kya state jaye, aur kaun kis ko check kare. |
Yeh terms Loop Engineering aur Harness Engineering mein dobara milenge. Yeh jaan boojh kar hai. Yeh course naqsha hai; woh courses ilaqa hain aur us cheez ko banate hain jis ka yeh sirf naam deta hai.
Tasveer: chaar containers, har aik agle ke andar
Containers, steps nahin. Har beat ab bhi prompt banata hai. Checker aur human gate loop par hain, beat ke andar nahin, kyun ke kaam banane wali cheez finished ka faisla nahin kar sakti.
Tasveer asal mein yeh jumla keh rahi hai. Khatam hone se pehle yeh teen baar aur ayega:
Bure context ke andar acha prompt fail hota hai. Bare harness ke andar acha context fail hota hai. Loop ke baghair acha harness khali baitha rehta hai.
"Har aik agle ke andar" ka amli matlab yehi hai. Inner layer missing outer ko nahin bacha sakti aur outer layer broken inner ko nahin bacha sakti.
90 seconds mein sab se aham claim ka saboot
Is course ki aik baat hoti dekh kar maanna aasaan hai. Apna pehle se use kiya agent kholein. Chota real task dein jis ka result check ho: chalne wali script, balance hone wala spreadsheet total, ya khulne wala link.
Kaam kahein, phir jawab parhein.
Isay theek karein taake yeh kaam kare, phir batayein kab complete hua.
Jawab par bharosa karne se pehle cheez khud check karein. Script chalayein. Column jorein. Link kholein.
Real work mein aksar confident "done" aise kaam ke saath milta hai jo verify hi nahin hua. Model ne jhoot nahin bola aur prompt kamzor nahin tha. Setup mein check lazmi nahin tha aur model ke ilawa kisi ne tay nahin kiya ke "done" kya hai. Woh ehsaas yaad rakhein. Concept 6 samjhata hai, aur baqi course is par khara hai.
15 minutes: oopar ki tasveer, Concept 2, aur Concept 9 ki table. Agli buri debugging session mein table saamne rakhein.
50 minutes: seedha poora, do short observation exercises aur end drill ke saath.
Baad mein: prose nahin, drill ke liye lautein. Recipes lookup ke liye hain, yaad karne ke liye nahin. Sirf aik jumla yaad rakhna hai, bilkul end par.
Part 1: Shakal
Concept 1: Containers, steps nahin
In chaar alfaaz ko aksar ladder banaya jata hai: neeche beginners ka prompt, oopar experts ka loop. Tasveer chupke se wada karti hai ke acha hone par lower steps peeche chhor denge.
Yeh tasveer ghalat hai, aur paise kharch karwane ke andaaz mein ghalat hai.
Har beat prompt banata hai. Chhe mahine se schedule, checker aur spine ke saath unattended loop bhi din mein kai baar model ko message bhejta hai. Message vague ho to loop vague kaam timer par tezi se banata hai. Inner layers peeche nahin chhorte; unhein wrap karte hain.
Rishta containment ka hai. Prompt window mein, window aik beat ke liye, aur beat ko run shuru, judge aur yaad rakhta hai.
Tasveer se teen ghalat yaqeen aasani se nikalte hain:
- Outer ka matlab later nahin. Inhein order mein build nahin karte. Aksar teen rent aur aik own karte hain, Concept 10 mein.
- Outer zyada important nahin. Khoobsurat loop ke andar careless prompt bad work ko schedule aur receipt ke saath banata hai. Layers ranked nahin, nested hain.
- Yeh aik box ke chaar sizes nahin. Yeh chaar mukhtalif objects hain, agla concept unhein alag karta hai.
Concept 2: Unit-of-work test
Har layer us kaam se define hoti hai jis ki woh zimmedar hai. Yeh chaar saaf ho jayein to naam kuch bhi ho, layer pehchan lenge.
| Layer | Unit of work | Aasaan alfaaz |
|---|---|---|
| Prompt | One model call | Jo type karke enter dabaya. Alfaaz badlein to prompt badal gaya. |
| Context | The window | Aik jawab ke waqt model jo dekh sakta hai: message, files, earlier turns, rules file aur tool results. |
| Harness | One beat | Agent aik jawab par nahin rukta. Tool call, result, soch, phir call. Instruction se quiet hone tak aik beat hai. Harness usay chalane wala code hai. |
| Loop | The whole run | Jab koi type nahin kar raha tab kya hota hai. Koi beat shuru, koi work judge, aur koi beats ke darmiyan yaad rakhta hai. |
Ab woh test jo vocabulary badalne par bhi chalega:
Blog, vendor, ya job description "harness", "context engineering", ya "agent loop" kahe to lafz par behas na karein. Poochhein: kis unit of work ki baat hai? One model call, window, one beat, ya poora run?
Un ke "harness" mein schedule ho to unhon ne is kitab ki do layers aik bana di hain. Tools aur credentials dene wala platform ho to aur wide ma'ni hai. Koi ghalat nahin. Maps alag hain. Ab confuse hone ke bajaye translate kar sakte hain.
Part 2: Chaaron layers aik aik karke
Concept 3: Prompt, unit one model call
Prompt aap ka banaya message hai: model ka role, aap ko kya chahiye, acha kaam kaisa hai, misaalein, aur jawab ki shakal. Aik input, aik response.
Yeh craft soch se choti hai. Sab se kamzor ingredient dhoondein aur sirf woh fix karein. Output shape ghalat ho to sahi example dein. Tone ghalat ho to audience batayein. Paanchon ko aik saath rewrite karne se sabab nahin pata chalta. AI Prompting in 2026 layer ko poora cover karta hai.
Log is par zyada invest karte hain kyun ke is ki practice code ke baghair aur edit das seconds mein hota hai. Donon achi baatein production problem mein trap hain: pressure mein log tootne wali cheez ke bajaye aasaan change pakarte hain.
Prompt waqai toota ho to: model task samjha aur qareeban sahi kaam kiya, magar jawab ghalat shape, length, voice mein ya assumed section ke baghair hai. Fact ghalat nahin. Apne alfaaz dobara parhein to wajah nazar aati hai.
Concept 4: Context, unit window
Window woh sab hai jo model aik response likhte waqt dekh sakta hai. Attach na hui file fact ke taur par available nahin. Kal ki conversation tab tak available nahin jab tak koi wapas na dale.
Lekin thin window se model khali nahin hota. Training knowledge gap bharti hai. Missing document par blank hone ke bajaye model sab se likely pehle se jani baat uthata aur truth ke confidence se kehta hai.
Yahi defining masla hai. Material hamesha fit se zyada hai, is liye koi chunta hai ke kya aur kis order mein jaye. Wohi curator hai. Aap likhein ya nahin, curator maujood hai aur window usay teen jobs deti hai.
Order pehle. Position material ka asar badalti hai. Mashhoor study mein important passage start ya end par ho to accuracy highest aur middle mein lower thi, "lost in the middle."1 Models 2023 ke the, is liye aaj effect test karein, assume nahin. Lesson baqi hai: attachment ke paragraph nine mein constraint dabana bhi aap ka decision hai.
Phir compression, jo free nahin. Chalis pages ko chaar mein summarize karke fit karte hain, magar dropped exception dobara nahin aata. Har compression bet hai ke kya matter nahin karega.
Aakhir dropping, jo policy hai. Window bharne par kuch bahar jayega. Policy aap na set karein to harness badtareen waqt par apni unread rule se karega.
Har setup par aik-minute test:
Window ke document par ishara karke batayein kis rule ne use rakha. Jawab "retriever ne return kiya" ho to curator nahin, search box hai. Kuch jobs mein theek, kuch mein dangerous. Give Your AI Searchable Context retrieval build karta hai. Agentic Coding, Parts 2 se 4, window haath se manage karna sikhata hai.
Context tootne ki shakal: jawab fluent, confident aur factually wrong. Aksar kisi aur cheez par sahi: old file, different customer, ya aap ke case ke bajaye docs example. Confident, wrong aur kisi sach ke qareeb hona signature hai.
Khud dekhein (2 minutes). Jaanay hue document par aik sawal do fresh chats mein karein. Pehli mein poora, doosri mein pehla third paste karein.
Yeh document hai. Sirf diye hue text par batayein ke [last third wali cheez] ke bare mein kya kehta hai?
Doosri chat aksar "pata nahin" ke bajaye confident jawab degi. Yeh bad behavior nahin; training knowledge window ka hole bhar rahi hai. Prompt badle baghair answer badal gaya.
Concept 5: Harness, unit one beat
Aap aik instruction dete hain. Agent teen files parhta, command chalata, error parhta, edit karta, phir command chala kar quiet ho jata hai. Aap ne aik baar type kiya, darjan cheezein huin. Yeh one beat hai aur harness usay chalane wala code.
Kaam aam hain: context assemble, model call, maange hue tools run, results feed, errors handle, aur beat end hone se pehle required proof enforce. Model ke rukne tak dobara.
Aap pehle se harness use karte hain. Claude Code, OpenCode aur Cowork mein hai. Permission prompt, startup rules file aur commit se pehle automatic check sab harness hain. Harness Engineering inherited default ke bajaye purpose se build karna sikhata hai.
Sub-agents tools ki tarah call aur bari cheez ki tarah behave hote hain. Tool result lautata hai. Sub-agent apni window kholta aur apna beat chalata hai. Tool-shaped opening se poore nested stack ka output aata hai.
Faida: chalis documents aap ki window se bahar parh kar teen paragraphs laut sakte hain. Nuqsan: summary full confidence se aati hai, ghalat hisse bhi. Aap ne source nahin dekhe, downstream bhi nahin dekhega. Confidence reading quality nahin batata.
Concept 6: Harness ko define karne wali limit
Beat timeout, token ceiling, error, denied permission ya model ke "done" par end ho sakta hai. Koi work success prove nahin karta. Sirf beat end prove hota hai.
Acha harness real checks enforce kar sakta hai: test suite, schema, known total. Command ke baghair finish refuse kar sakta hai. Yeh verification hai aur jahan mumkin ho build karein.
Magar boundary aham hai:
Beat specific check pass hona prove kar sakta hai. Yeh decide nahin kar sakta ke woh check kafi tha.
Passing test sirf test pass prove karta hai, coverage ya poore task ki definition nahin. Run se pehle kisi ko "finished" define karna hai.
Bank-reconciliation agent 40 matches aur balanced statement report karta hai, magar totals compare nahin karta. Context aur tools theek, errors zero. Claim phir bhi false.
Checker tab pakarta jab kisi ne totals matching ko pehle required kiya hota. Maker apni finish line bana kar khud certify nahin kar sakta.
Prompt mein "verify" likhna hal nahin. Model "verified" bhi "done" ki tarah likh sakta hai. Final success definition judged work ke bahar se aati hai, is liye agle outer layer ki hai.
Concept 7: Loop, unit poora run
Loop woh deta hai jo beat khud nahin de sakta.
Heartbeat beat shuru karta hai: schedule, event, ya condition. Is ke baghair aap heartbeat hain; type band to kaam band.
Spine model se bahar state rakhta hai taake agla beat pichla kaam jane:
run: nightly reconciliation, 2026-03-14
done: pulled 412 payments and 388 open invoices
in progress: matching pass 3 of 5, 341 matched so far
needs a person: invoice 4471, two candidates both at 0.52
budget: 3 beats used of 12
File clever nahin, isi liye chalti hai. Agla beat isay parh kar pass 3 se resume karta hai. Chauthi line agent ka undecided case hai; Concept 8 isay human gate banata hai.
Outside stops poore run ka continuation tay karte hain:
- Advance mein chuni aur command se proved success condition. Criterion maker ke bahar se.
- Beats, spending aur elapsed time limits. Ceiling promising attempt ki parwah nahin karti.
- No-progress check. Meaningful change ke baghair repetition pakarta hai.
- Alag checker. Judge ne work nahin banaya.
Koi maker se nahin poochta ke finished hai. Har aik maker se bahar established fact par hai. Yehi maker-checker rule hai.
Opening failure mein missing no-progress check tha, better prompt nahin. Prompt words badal sakta, run ko stuck notice nahin kara sakta.
Loop Engineering heartbeat, spine, stops, checker aur gate banata hai. Trusting the Checker checker par trust poochta hai.
Concept 8: Human gate exit hai, stop nahin
Stops run ko mehfooz fail karte hain. Gate run ko madad ke saath succeed karata hai.
Asal baat: ambiguity error nahin hai. Do payments 0.52 hon to matcher malfunction nahin, data aisa hai. Gate ke baghair fail karke nau passes phenkna ya guess karna hai.
Guess zyada khatarnak hai. Crash loud hai. Guess same format, confidence aur report position mein correct answer jaisa hai. Downstream farq nahin dekh sakta. Failed run fix hota hai; guessed run spreadsheet ka chupchaap ghalat number hai.
Gate advance mein likha jata hai: confidence line, value limit, ya hard-to-undo action. Named person, maslan accounts payable ki Ayesha, donon candidates dekh kar faisla karti hai.
Jawab run end nahin karta. New evidence ke taur par re-enter hota aur run pause se jari rehta hai.
Part 3: Naqsha use karna
Concept 9: Kaun si layer tooti?
| Jo nazar aaye | Pehle kahan dekhein | Kya badlein |
|---|---|---|
| Shape ya tone ghalat, task samjha | Prompt | Weakest ingredient: examples, instructions, output shape |
| Confident, fluent, factually wrong | Context | Curator: kya aya, order, kya drop hua |
| Unproved success ya unnoticed failed tool | Harness | Tools, error handling, required proof |
| Wrong answers unchecked pohanche | Loop | Checker, criterion owner, kya kabhi fail kiya |
| Rukta nahin, jaldi rukta, ya poochne ke bajaye guess | Loop | Stops aur gate |
Yeh search order hai, verdict nahin. Failures boundaries cross karti hain. "Pehle yahan" parhein, "blame yahan" nahin.
Table easy-edit habit todti hai. $50 repeat par system prompt rewrite hota hai jab no-progress check missing tha. Pressure mein easy layer pakarti hai. Kuch chhoone se pehle layer ka naam buland kehna discipline hai.
Concept 10: Aap kaun si layers own karte hain?
| Layer | Mode 1: general agent se problem solve | Mode 2: worker manufacture |
|---|---|---|
| Prompt | Zyada tar aap ka; platform system instructions own karta hai. | Aap ka, aik baar likha aur reused. |
| Context | Kuch aap ka; compaction aur retrieval harness ke. | Aap ka; curator aap likhte hain. |
| Harness | Rented, tools, skills, hooks se configurable. | Aap ka. |
| Loop | Mostly rented; platform caps aur approvals de sakta hai. | Aap ka; har stop aap ka. |
Mode 1 mein no-progress file nahin. Outer job rented behavior janna hai: compaction kya phenkta, retries, aur room khatam hone par kya hota hai.
Mode 2 mein chaaron aap ki hain.
Mode 1 mein kehna, "mera tool yeh sab karta hai, is liye mein bhi karta hun."
Tool apne kaam ke liye karta hai. Aap ke worker ke liye nahin. Claude Code apne beats cap karta, aap ka loop nahin. Isi gap mein demo incident banta hai.
Kya rent kiya (4 minutes). Agent chun kar pehle guesses likhein:
- Window bhare to kya hatta aur kya batata hai?
- Tool fail ho to retries kitni aur visible hain?
- Mid-task limit par completed kaam ka kya hota hai?
Docs dekhein aur long session se compaction test karein: teen early decisions agree, compaction tak kaam, phir restate. Jo na aaye harness ne unnecessary maana.
Answers team-readable jagah likhein. Mode 1 mein woh chota document aap ka context aur loop work hai, chota version nahin balki mukhtalif job.
Concept 11: Graphs kahan fit hote hain
Graph engineering mashhoor hai magar graph fifth layer nahin.
Chaar layers aik execution path hain: message, window, beat, run. Graph topology hai: agla kya chale, edge par kya move ho, aur kaun kis ko check kare.
Chaar layers node ke andar ka waqia batati hain. Graph nodes ke darmiyan ka waqia batata hai.
Node function, rule, tool call, human gate, measurement, beat, loop ya agent ho sakta hai. Multi-agent sirf aik qisam hai.
- Execution graph next step aur state movement tay karta hai.
- Memory graph entities, findings aur sources later runs ke liye rakhta hai.
- Governance graph record karta hai kaun feed, check, approve aur constrain karta hai.
Yeh kitab ke labels hain, standard industry vocabulary nahin.
Nightly accounts-payable graph mein chhe nodes aur aik agent:
- Pull. Function payments aur invoices common format mein lata hai.
- Route. Rule panch lakh rupay se baray invoices human approval mein bhejta hai.
- Match. Full agent matches aur confidence scores deta hai.
- Gate. Ayesha low-confidence/high-value cases decide karti hai; jawab Match mein evidence banta hai.
- Prove. Function matched aur statement totals compare karta; mismatch rokta hai.
- Post. Function accepted matches likhta aur unresolved review ko bhejta hai.
Do conditions Gate ko bhejti aur decision Match ko lautta hai. Har edge payload ka naam rakhta hai.
Sirf Match agent hai. Baaqi deterministic code, rule, person aur measurement. Match ke andar chaar layers hain; graph us se pehle aur baad ka sawal hai.
Prove Match se bahar aur unskippable hai. Maker judging measurement ko overrule nahin kar sakta. Edges bhi aham: normalized records, proposed matches, confidence aur human decision. Unclear edge contracts graph debugging ko mushkil karte hain.
Cost zyada tar Match mein hai. Paanch nodes qareeban free hain. Boxes nahin, agentic nodes aur frequency measure karein.
June 2025 mein Anthropic ne single-agent research ke liye chat se qareeban 4x aur multi-agent ke liye 15x tokens report kiye.2 Yeh dated system measurements hain, constants nahin.
Shared context aur strong dependency multi-agent ka poor fit hai.3 Tools badle, constraint nahin. Graph Engineering build aur not-build dono sikhata hai.
Concept 12: Jab framework aap se larta hai
Kuch failures two-layer hain. Old turns drop karna context hai, magar long loop window bharta hai. Donon toot sakte hain; cheaper fix system par hai.
Diagnosis aik layer, fix doosri. Unproved success diagnosis harness limit, fix outer chosen criterion. Framework ne symptom ke ilawa jagah bhej kar kaam kiya.
Kuch chaaron mein nahin. Model quality nahin de sakta to containers capability nahin banate. Better model, chota task, ya different approach. Trusting the Checker pata lagata hai.
Har project ko chaaron nahin chahiye. One-off ke liye loop waste ho sakta hai. Map existence batata hai, obligation nahin.
Part 4: Practice
Drill: layer ka naam lein
Har failure par pehli layer aur aik change likhein. Key se pehle jawab dein.
1. Five-column table maangi, teen accurate prose paragraphs aaye.
Answer
Prompt. Kaam sahi, sirf shape ghalat. Desired table ka example dein.
2. Pricing cap 5,000 bataya, page 50,000 hai aur agent ne page parha.
Answer
Context. Confident, fluent, document par wrong. Compression, truncation ya old copy check karein.
3. Overnight "tests passing" magar suite chali nahin aur build red.
Answer
Diagnosis harness, fix loop. Hook suite run aur beat ko enforce kare. Trustworthy outer stop advance chosen success criterion hai.
4. Run budget kharch karke teen approaches repeat karta hai.
Answer
Loop. No-progress check aur spending limit missing.
5. 300 invoices matched; ambiguous cases mein silent guess.
Answer
Loop, missing gate. Low-confidence/high-value trigger person ko route kare.
6. Sub-agent 40 tickets ki summary mein do wrong claims lata hai.
Answer
Harness. Quotes, ticket IDs aur checkable receipts required karein.
7. Long session early agreed decision bhool kar contradict karta hai.
Answer
Context. Window se decision drop hua. Rules file, spec ya reattached note mein durable karein.
8. Chaaron layers clear, hard research output phir mediocre.
Answer
Koi nahin. Capability issue: stronger model, smaller task, different approach ya person.
Ab apne failure par karein
Pichla wrong ya expensive agent output soch kar order mein:
- Kaun si unit ghalat hui? Call, window, beat ya run?
- Baad mein kya badla? Same layer?
- Kya pakarta? Mechanism, layer, criterion owner.
Misaal: legal team auto-renew contracts dhoondti hai. Agent 40 reviewed aur 3 renewals kehta hai. Do mahine baad 3 aur "evergreen term" se renew hote hain.
1. Unit? Shape aur documents theek. Run ne "reviewed" ko phrase search bana kar khud judge kiya. Whole run failed.
2. Change? Prompt mein evergreen, rolling term, self-renewing jore. Known examples fix, next wording nahin. Layer 4 failure ko layer 1 par badla.
3. Catch? Legal lead ki advance condition:
- har contract quoted clause aur page number de; ya
- no renewal clause found de aur person ko jaye.
Command 40 complete nonblank results check kare. Agent rule weak nahin kar sakta. Missing contract silent success ke bajaye visible question hota.
Useful skill instant label nahin; tempting easy layer par rukne se inkar hai.
Mera recent agent failure: [ask, output, aur ghalati ka pata]. Chaar layers mein broken layer dhoond raha hun: prompt (call), context (visible facts), harness (beat aur tools), loop (run, start, stop, check). Narrow karne ke sawal poochhein, best guess aur earlier catch batayein. Meri layer ghalat ho to push back karein.
Real failures drill jitne clean nahin. Aik real case aath answers pehchanne se zyada sikhata hai.
Yahan se kya le jayein
- Concept 1. Nested containers, ladder nahin; har beat prompt banata hai.
- Concept 2. Units: call, window, beat, run.
- Concept 3. Weakest prompt ingredient fix karein; easy layer ko doosre bugs ka blame milta hai.
- Concept 4. Window facts aur training holes bharti hai; curator order, compress, drop karta hai.
- Concept 5. Beat instruction se quiet tak; sub-agents nested stack aur unearned confidence lautate hain.
- Concept 6. Beat check prove karta, enough nahin; maker finish judge nahin karta.
- Concept 7. Loop heartbeat, spine aur outside stops deta hai.
- Concept 8. Gate exit hai; ambiguity error nahin, silent guess crash se khatarnak.
- Concept 9. Fix se pehle layer ka naam.
- Concept 10. Mode 1 mein rent; Mode 2 mein own. Tool ka kaam worker ka nahin.
- Concept 11. Graph nodes ke darmiyan topology, fifth layer nahin.
- Concept 12. Map search order hai; capability create nahin karta.
Yaad rakhne wala jumla:
Bure context ke andar acha prompt fail hota hai. Bare harness ke andar acha context fail hota hai. Loop ke baghair acha harness khali rehta hai. Kuch tootay to fix se pehle layer ka naam lein.
Demo inner layers se chalta, production outer layers maangta hai. Masla model nahin, missing layers hain.
Agla raasta
- Harness Engineering layer 3 build karta hai.
- Loop Engineering layer 4: heartbeat, spine, maker-checker, gate.
- Trusting the Checker checker ka trust.
- Graph Engineering one path ke baad topology.
- Leaving the Laptop trusted loop kahan rahe.
Layers 1 aur 2: AI Prompting in 2026 aur Agentic Coding, Parts 2 se 4.
Kaam karne wale agents clever prompts se nahin, aise loops se bante hain jo jante hain kab rukna aur kab poochna hai.