OpenAI Ki Agents API: Public Beta Actually Kya Change Karti Hai
Beta Status, October 2026
Ye post October 2026 mein refresh hui hai aur Agents API abhi bhi public beta ke tor par describe hoti hai — ye neeche har cheez ke liye honest caveat hai. Beta interfaces, headers, aur key scopes bina kisi launch post ke shor ke change ho sakte hain, is liye yahan quote ki gayi kisi bhi cheez par build karne se pehle current official API reference dobara check karein. Neeche kuch bhi ek stable, generally-available contract ke tor par nahi parhna chahiye.
OpenAI ne apni Agents API ka public beta September 10, 2026 ko khola, apni khud ki developer post aur X par ek thread ke zariye announce kiya gaya. Headline claim straightforward hai lekin undersell karna aasan hai: wahi harness aur infrastructure jo internally Codex chalata hai ab kisi bhi developer ke liye ek general-purpose API ke through available hai, sirf OpenAI ki apni product team ke liye nahi.
Actually Kya Launch Hua
API chaar primitives ke around organize hoti hai. Ek agent ek model, uski instructions, uske available tools, aur kisi bhi connected MCP servers ka combination hai. Aur ye tools hi sab se fragile hissa sabit ho rahe hain — ek nayi research ke mutabiq silent tool errors e-commerce workflows mein agent ki kamyabi lagbhag aadhi kar dete hain. Ek environment ek optional sandbox hai jo file access aur command execution deta hai, taake ek agent actually cheezein kar sake, sirf unke baare mein text generate na kare. Ek session ek durable instance hai jo multiple turns ke across state persist karta hai, taake ek multi-step task har baar scratch se manually re-assemble na karna pade. Events aur items simply wo inputs hain jo agent ko bheje jate hain aur wo outputs jo wo return karta hai.
Un logon ke liye jinki aankhein “primitives” par glaze ho jati hain, hamara khud ka plain-language explanation ke API actually kya hoti hai us foundational concept ko cover karta hai jis par ye banti hai — ek API ek software ke liye doosre se kuch maangne aur ek predictable jawab wapas paane ka defined tareeqa hai. Ye wahi idea hai, scaled up ek poore agent ke ongoing behavior ko cover karne ke liye, ek single request ke bajaye.
Agents API vs Agents SDK vs Responses API
Naming hi sab se zyada confuse karti hai, to farq seedha samajh lein. Agents API hosted harness hai: session, sandbox aur recovery OpenAI chalata hai, aur aap ise HTTP par call karte hain. Agents SDK code-first framework hai: agent aapki apni application mein banta hai, aur hosting aur state aapki apni zimmedari hoti hai. Responses API single-request model call hai: powerful zaroor, lekin har call stateless hoti hai — conversation state aap khud assemble karte hain. Zapier, Make aur n8n in teenon se bahar hain: wo existing apps ko triggers aur actions se jorte hain, aur kabhi koi agent harness run nahi karte.
| Surface | Harness kaun chalata hai | State kaise handle hoti hai | Kab choose karein |
|---|---|---|---|
| Agents API | OpenAI (hosted) | Sessions multiple turns ke across persist karti hain | Jab aap ek agent product bana rahe hon aur Codex infrastructure chahiye ho |
| Agents SDK | Aap, apne code mein | Aap khud manage karte hain | Jab aapko apne stack ke andar framework control chahiye ho |
| Responses API | Koi nahi — ek waqt mein ek call | Aap har call par khud assemble karte hain | Single model calls ke liye, ongoing agent kaam ke liye nahi |
| Zapier / Make / n8n | Automation platform | Trigger → action chains | Existing apps ko aapas mein jorne ke liye, agents banane ke liye nahi |
Pehli Session, Ek Hi Call Mein
Pehli call ki shape, session banane ke liye OpenAI ki apni API reference se:
curl https://api.openai.com/v1/agents/sessions \
-H 'Content-Type: application/json' \
-H 'OpenAI-Beta: agents=v1' \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-d '{"environment": {"type": "none"}}'
Teen cheezein note karne laiq. Pehli: raw HTTP clients ko OpenAI-Beta: agents=v1 header khud bhejna parta hai; official SDKs ise automatically laga dete hain. Doosri: official quickstart ke mutabiq session operations ke liye API key scopes api.agents.read aur api.agents.write chahiye, aur model inference ke liye api.responses.write — key ko sirf us kaam tak scope karein, aur ise apne server par rakhein, kabhi browser code mein ya agent sandbox ke andar nahi. Teesri: ek session ka available rehna is baat ka saboot nahi ke uska latest task kamyab hua — sirf status nahi, tool calls aur final result parhein.
Ye “ek aur API” se Kyun Bara Deal Hai
Asal news yahan ek naya endpoint nahi hai — ye OpenAI ka wo infrastructure package karna hai jo usne apne internal Codex product ke liye banaya (session persistence, context compaction jab ek long-running task apna available context bhar de, ek failure ke baad automatic recovery) aur ise ek general-purpose product banana jise koi bhi developer build kar sake. Us tarah ki infrastructure scratch se banana genuinely hard engineering work hai jo zyada tar chhoti teams kabhi properly karne tak nahi pahunchti, yehi exact wajah hai ke is point se pehle zyada tar “AI agent” products ya to ek single API call ke around thin wrappers the ya unhe ek team chahiye thi jo real session aur sandbox infrastructure in-house banane jitni bari ho.
Developers Actually Agent Kahan Chalate Hain
Ek developer agent ka compute ek OpenAI-managed sandbox, apni khud ki infrastructure, ya ek partner sandbox mein chala sakta hai — launch ki coverage ne specifically Cloudflare, DigitalOcean, aur Oracle ko sandbox environments offer karne wale partners ki tarah naam liya. Ye flexibility exactly us tarah ke concern ke liye matter karti hai jo ek chhoti agency ya team rakhti — ek managed sandbox kuch ship karne ka fastest path hai, jab ke existing infrastructure par chalana zyada matter karta hai ek baar data residency ya cost control ek real consideration ban jaye. Aur safety story sandbox choice jitni hi eham hai — is haftay ke rogue agent waqiyat batate hain ke production touch karne se pehle kaunse guardrails lagane chahiye.
Ye Zapier, Make, Ya n8n Se Kaise Alag Hai
Ye explicitly kehna worth hai ke ye kis level par operate karta hai, kyunke ise un no-code automation tools ke sath confuse karna aasan hai jo already is site par cover ho chuke hain. Zapier, Make, aur n8n existing apps ko triggers aur actions ke through connect karte hain — zyada tar chhoti ecommerce ya agency automation needs ke liye genuinely sahi tool. Agents API us se ek level neeche hai: ek developer-facing primitive un teams ke liye jo actually ek custom agent product bana rahi hain, existing tools ko wire karne ke bajaye. Ye padhne wale zyada tar chhote businesses ko directly Agents API ki zaroorat nahi — unhe Zapier ya Make chahiye, ya ek developer jise wo hire karein jo neeche is jaisi kisi cheez use kar sakta hai.
Early Numbers, Coverage Ko Attributed, Independently Verified Nahi
Launch ki coverage ne early-adopter results cite kiye — SafetyKit ki reported 60% cost reduction, Hypha ki reported 86% fewer failures, aur Cirridae ki reported 4x faster latency. Ye third-party claims hain jo launch coverage ne relay ki hain, wo numbers nahi jo is site ne independently verify kiye hon, aur inhe un companies ke early, self-reported results ki tarah padhna chahiye jinka ek successful launch narrative mein interest hai, ek neutral benchmark ke bajaye.
Koi Extra Fee Nahi, Lekin Actually Free Bhi Nahi
Harness khud koi separate charge nahi rakhti — usage existing model token costs, tool usage, aur sandbox compute minutes ke through bill hoti hai. Ye kisi ke liye bhi matter karta hai jo cost estimate kar raha ho: API access ka ek separate fee se “free” hona iska matlab nahi ke ek real agent product ise chalana sasta hai, kyunke underlying token aur compute costs abhi bhi apply hoti hain aur ek long-running, tool-heavy agent ke liye jaldi jama ho sakti hain.
Launch Ke Baad Kya Naya Aaya: Computer Use (DevDay, 29 September)
Beta September 10 ke baad ruka nahi. DevDay par, 29 September ko, OpenAI ne computer-use tool add kiya: hosted browser ke zariye software chalane wale agents, aur browser sign-ins jaise sensitive actions ke liye approval requests. Is se API “code chalane wala agent” se “wo tools use karne wala agent jo aapki team pehle se use karti hai” ban jati hai — aur risk bhi usi ke sath move karta hai. Koi bhi workflow jo agent ko aapki tarah sign in karne de, us mein approval step aur pehli runs par ek insaan ki nazar honi chahiye, sirf ek success metric nahi.
Banane Se Pehle Jaanne Laiq Limits
Chaar decision-blockers jinhain launch coverage aksar chhupa deti hai. Data residency beta mein sirf US hai. Zero Data Retention available nahi hai, self-hosted environments ke liye bhi nahi — regulated teams ise roadmap item nahi, hard constraint samjhein. Multi-agent runs mein concurrent subagents ki cap hai (reported default 6), to fan-out designs mein sirf enthusiasm nahi, budgeting chahiye. Aur ye beta hai: interfaces, upar wala beta header samet, badal sakte hain, to integration ko current reference par pin karein aur kisi production commitment se pehle ise dobara check karein.
Ek Safe Pehla Pilot
Naye agent ko pehle din production tools mat dein. Ek: ek bounded, read-only kaam chunein jis ka success rule shuru karne se pehle likha hua ho. Do: ise pehle read-only tools ke sath chalao aur poora run review karo — har tool call, sirf final answer nahi. Teen: ise mushkil test cases dein (ghalat input, missing files, ambiguous instructions) aur dekhein jab ye kaam poora na kar sake to ye kya karta hai. Chaar: sirf tab permissions barhayein, ek tool ek waqt mein, approval flow on rakhte hue. Jo teams jalti hain wo wo nahi hain jin ke agents zor se fail hote hain; wo wo hain jin ke agents khamoshi se ghalat cheez mein kamyab ho jate hain.
Ek Chhoti Team Ke Liye Iska Abhi Kya Matlab Hai
Realistic takeaway “is hafte ek agent bana lein” nahi hai — ye pehchanna hai ke ye actually kab apply hota hai. Agar goal ek internal workflow automate karna hai (order confirmations, ek customer support draft, ek reporting pipeline), hamara existing automation tools guide abhi bhi sahi starting point cover karta hai. Ye API tab relevant ban jati hai jab goal ek internal process automate karne se shift ho kar ek actual agent-based product banane ki taraf jata hai jise ship ya sell karna maqsad hai, jahan underlying session aur sandbox infrastructure warna scratch se banana parta.
OpenAI Ke Is Saal Ke Ek Broader Pattern Ka Hissa
Ye release wahi pattern follow karti hai jo hamare GPT-6 Astra explainer mein cover hua — OpenAI ek fast pace par developer-facing infrastructure ship karta raha hai, har release ko us basis par evaluate karne laiq hai ke ye actually kya change karti hai, ye assume karne ke bajaye ke har announcement equal weight rakhta hai. Har model ya API release ek small business audience ke liye equally matter nahi karti; ye specifically un logon ke liye matter karti hai jo software bana rahe hain, ek store owner ke liye nahi jo agli automation shortcut dhoond raha ho.
FAQ
Kya mujhe Agents API chahiye, ya Zapier/Make kaafi hai?
Zyada tar chhote businesses ke liye Zapier ya Make. Wo existing apps ko triggers aur actions se jorte hain, jo order confirmations, support drafts aur reporting pipelines ko cover karta hai. Agents API sirf tab chahiye jab aap ek actual agent product bana rahe hon — software jo aap ship ya sell karein — jahan session persistence aur sandbox infrastructure warna scratch se banana parta.
Agents API ki billing kaise hoti hai?
Harness ka koi separate fee nahi hai. Aap model token costs, tool usage aur sandbox compute minutes ke liye pay karte hain. Ek long-running, tool-heavy agent is liye mehnga par sakta hai chahe “API” free ho — budget tokens aur compute par banayein, API calls par nahi.
Kya Agents API generally available hai?
Nahi — ye public beta hai. Interface changes expect karein, beta header aur key scopes ko official reference ke sath current rakhein, aur beta behavior par bina fallback ke koi business-critical dependency mat banayein.