Aapka AI Agent Kharab Nahi — Uski Tools Chup Chaap Jhooth Bol Rahi Hain

21 October ko HKUST aur Tencent ke researchers ne RITE benchmark release kiya jo AI agent failures ke baare mein aapki soch badal sakta hai. Unhone nine e-commerce workflows se 167 matched units liye aur ek shararti kaam kiya: agent ki tools se plausible-lekin-ghalat values wapas dilwayin — koi error message nahi, koi red flag nahi, sirf confident ghalat data. Natija: agent ki kamyabi ki sharah lagbhag aadhi reh gayi.

Is liye nahi ke model bewakoof ho gaya tha. Is liye ke kisi ne use bataya hi nahi ke data ghalat tha.

Ye wo failure mode hai jis par koi baat nahi karta. Hum prompts, context windows aur kaunsa model use karein — in par obsessed rehte hain, jabke asal toot-phoot ek layer neeche hoti hai: tool outputs mein, jin par agent aankh band karke bharosa karta hai. Hum real stores mein agents lagane mein itna waqt guzaar chuke hain ke is zakhm ka nishan hamare paas bhi hai.

$3.50 wali pricing ghalti jise kisi log ne nahi pakra

Pichhli quarter humne ek client store ke liye repricing automation chalayi thi. Koi exotic cheez nahi — ek agent jo har subah supplier ki price feed check karta aur 35% margin hold karne ke liye 40 product listings adjust karta. Chhe haftay tak bilkul saaf chali.

Phir supplier ki feed ne cached page serve kar diya. Unke best-selling SKU par $18.49 wholesale cost $14.99 wapas aayi — pichhle mahine ki price, 200 OK aur fresh timestamp ke sath. Agent ne exactly wahi kiya jo use kaha gaya tha: margins recalculate kiye, retail price $28.49 se $23.99 kar di, aur listing update kar di. Kisi ko pata chalne se pehle gyarah orders aa gaye. Store ko har unit par lagbhag $3.50 ka nuksan hua — kul taqreeban $38 — aur agent ke logs mein perfect run nazar aayi. Koi error nahi. Koi warning nahi. Sirf ek tool jisne confident hokar ghalat jawab diya.

Ye hai silent tool error. Hallucination nahi — model ne kuch invent nahi kiya. Crash nahi — kuch fail nahi hua. Ek tool ne ghalat data wapas kiya, achhe data ke kapron mein, aur agent ke paas shak karne ki koi wajah nahi thi.

Silent errors asal duniya mein kaise dikhte hain

Jab aapko iski shape pata chal jaye, to ye har jagah nazar aata hai:

• Stale inventory counts. Aapke warehouse API ke mutabiq 14 units stock mein hain; asal number 2 hai kyunke sync job Jumme ko mar gayi thi. Agent teen orders confirm kar deta hai jo ship hi nahi ho sakte. Cancellations aur buri reviews aapke hisse mein.
• Cached shipping quotes. Rate tool cached response se $6.90 wapas karti hai; asal carrier rate pichhle haftay $14.20 ho gaya tha. Agent checkout par customers ko purana number quote karta hai. Kisi ko pata chalne tak har order par $7.30 aapki jeb se.
• Ghalat-record lookups. “Maria Garcia” ki CRM search ghalat Maria Garcia wapas karti hai — same naam, mukhtalif customer. Agent order update (kisi aur ke items ke sath) ghalat inbox mein bhej deta hai. Ab ye privacy ka masla hai, sirf bug nahi.

Sab se common ghalti jo hum dekhte hain, wo ye hai ke teams inhein model ka masla samajhti hain. Agent ne ghalat kaam kiya, to system prompt rewrite kar diya. Ya newest model par upgrade kar liya, jaise zyada smart dimagh jhoothi API pakar leta. Nahi pakarta. Behtar model sirf zyada articulate victim hai — wo bohat convincingly explain karega ke usne ghalat number par bharosa kyun kiya.

Teen jagah jahan agents actually fail hote hain

RITE ke researchers failure ko teen successive bottlenecks mein todte hain, aur agents banane wale kisi ke liye ye useful mental model hai:

1. Recognition. Kya agent ne notice bhi kiya ke kuch garbar hai? Zyada tar nahi karte. $14.99 model ko ghalat nahi lagta — ye bilkul reasonable number hai. Koi “ye stale lag raha hai” wali instinct nahi hoti jab tak aap khud na banayein.
2. Refetch. Agar use shak ho bhi jaye, to kya wo dobara try karta hai? Ek retry aksar transient glitches fix kar deti hai — lekin silent errors transient nahi hote. Cache wahi ghalat number dobara serve karega. Usi tooti source se dobara poochna jhoothe se dobara jhooth bulwane jaisa hai.
3. Integration. Agar agent ko corrected data mil bhi jaye, to kya wo use waqai apne plan mein wapas fold karta hai — ya pehle dekhe number par hi kaam karta rehta hai? Humne agents ko dekha hai jo ek message mein correction acknowledge karte hain aur teen steps baad original value par act karte hain.

Zyada tar agent frameworks — jin mein OpenAI ka Agents API bhi shamil hai, jo ab public beta mein hai — aapko ye checks banane ke tools dete hain. Taqreeban koi banata nahi, kyunke demo mein kabhi jhoothi API shamil nahi hoti.

Paanch defenses jo waqai kaam karti hain

Hum aapko “zyada validation add karein” kehkar nahi chhorein ge. Ye raha wo jo hum actually implement karein ge, leverage ke order mein:

1. Paise wale har kaam se pehle verify karein. Agent ko kabhi single tool read par charge, refund, reorder ya reprice nahi karna chahiye. Hamara rule: koi bhi action jo paisa move kare ya customer-facing price change kare, usay do independent reads chahiyein jo agree karein. Usi cached feed ke do reads count nahi hote — wo usi jhoothe se dobara poochna hai. Ek read supplier feed se, ek aapke apne database se, aur dono ka match hona zaroori hai.

2. Critical reads par refetch-and-compare, tripwire ke sath. Inventory aur pricing ke liye do baar parhein, 30 second ke gap se. Agar values disagree karein, to rukein aur human ko escalate karein. Ye transient glitches pakarta hai (sasta fix) aur systematic wale surface karta hai (wo cache jo kabhi update nahi hota).

3. Prices aur stock ke liye second source of truth rakhein. Aapke store ka database authority hai; supplier feed suggestion hai. Jab dono mein aapke threshold se zyada farq ho — hum price par 5%, stock par 10% use karte hain — to agent pause kare aur flag kare. Disagreement information hai. Khamoshi dushman hai.

4. Har tool output ko timestamp ke sath log karein. Sab ko. Post-incident reviews mein sab se common ghalti jo hum dekhte hain, wo teams hain jinke paas zero tool-output logs hain. Unke paas agent ki reasoning hai (“maine price $23.99 update karne ka faisla kiya”) lekin wo data nahi jis par usne reasoning ki. Tool outputs ke baghair aap bata nahi sakte ke model ghalat tha ya data ghalat tha — aur ye bilkul mukhtalif fixes hain. Storage sasti hai. Pachhtawa mehnga hai.

5. Autonomy par dollar threshold lagayein. $200 se upar ke impact wala koi bhi kaam agent karna chahe to human ka tap zaroori hai. Is se neeche wo chalta rahe. Ye agent par distrust nahi hai — ye blast radius size karna hai taake silent error aapko hazaaron ka nahi, darjanon dollars ka pare. Businesses ke liye agent guardrails wali hamari playbook iska permission-scoping wala pehlu cover karti hai; dollar threshold uska financial twin hai.

Kya kaam nahi karta

Buri practices ka naam lete hain, kyunke ye bikti rehti hain:

• “Vendor data validate karta hai.” Nahi. Aapke supplier ki API, aapka shipping plugin, aapka CRM — in mein se koi ye promise nahi karta ke data current hai, sirf ye ke request kamyab hui. HTTP 200 delivery receipt hai, truth certificate nahi.
• Prompt mein fix karna. System prompt mein “act karne se pehle sab prices double-check karo” likhna fix jaisa lagta hai. Nahi hai. Model us number ko double-check nahi kar sakta jis ka source uske paas hai hi nahi. Prompt engineering ground truth conjure nahi kar sakti.
• Bare models. Usi jhoothi tools ke sath zyada capable model zyada eloquently fail hota hai. Hum pehle bhi keh chuke hain; kehte rahein ge jab tak benchmark charts ghalat parhe jate rahein ge.

Andar se ye har tool ek API call hai — doosre software se ek defined request, structured jawab ke sath. Agar ye mental model fuzzy hai, to hamara API ka plain-language explainer das minute ka mustahiq hai, kyunke is article ki har cheez usi samajh se neeche ki taraf behti hai — ke API actually kya promise karti hai (aur kya nahi).

Ek imandar caveat

RITE ek achha benchmark hai, lekin synthetic hai: researchers ne exactly chuna ke har tool kab aur kaise jhooth bole. Asal duniya ke silent errors zyada messy hain — caches ajeeb intervals par expire hote hain, sync jobs Jumme ko marti hain, insaan SKUs ghalat type karte hain. Humein production agent deployments mein silent errors ki asal per-call rate nahi maloom, aur jo bhi aapko ek number quote kare wo andaza laga raha hai. Benchmark specifically e-commerce workflows cover karta hai; “lagbhag aadhi” wali figure code-review agents ya research agents par hold karti hai ya nahi, ye open question hai. Humein abhi yaqeen nahi.

Jis par humein yaqeen hai: failure mode real hai, ye under-instrumented hai, aur fix boring engineering hai — redundant reads, logged outputs, human thresholds — zyada smart model nahi. Jo teams ye internalize kar lein gi, unke agents real APIs se takrane par survive karein ge. Jo nahi karein gi, wo apne prompts debug karti rahein gi jabke unki tools seedhi shakal banakar unse jhooth bolti rahein gi.

FAQ

Silent tool error exactly kya hota hai?

Kisi tool ka koi error signal diye baghair plausible-lekin-ghalat value wapas karna — masalan ek cached price jo 200 OK aur fresh timestamp ke sath serve ho. Ye hallucination nahi hai (model ne kuch invent nahi kiya) aur crash bhi nahi (kuch fail nahi hua). Agent ke paas data par shak karne ki koi wajah nahi hoti, is liye wo us par confidently action leta hai.

Zyada smart model isay kyun theek nahi karega?

Kyunke failure model se ek layer neeche hai, un tool outputs mein jin par wo bharosa karta hai. Usi jhoothi tools ke sath zyada capable model sirf zyada eloquently fail hoga — wo bohat convincingly explain karega ke usne ghalat number par bharosa kyun kiya. Prompt engineering se ground truth paida nahi hoti.

Sab se zyada leverage wali defense kaunsi hai?

Paise se mutaliq koi cheez move hone se pehle do independent reads jo agree karein — supplier feed se ek read plus apne database se ek read, na ke usi cached feed ki do reads. Wahi jhootha do baar, wo count nahi hota.

Kya RITE benchmark ye sabit karta hai ke production mein bhi yehi rate hai?

Nahi, aur post is baare mein imaandar hai: RITE synthetic hai, researchers ne exactly chuna ke har tool kab aur kaise jhooth bole. Asal duniya ke silent errors zyada messy hain, aur production deployments mein asal per-call rate koi nahi janta. Failure mode real hai; exact rate settled nahi hai.