OpenAI Opens Its Agents API to All Developers, Putting the…

The managed API absorbs session state, context compaction, and subagent coordination — the plumbing that has slowed down every team trying to ship a small-business sales or outreach agent.

OpenAI put its Agents API into public beta on Thursday, September 10, exposing the Codex harness that runs ChatGPT for Work behind a single API call. The company frames the launch as the productized output of scaling that harness “to millions of people,” and the practical read is simpler: the hardest infrastructure problems in agent development, session state, context compaction, crash recovery, subagent orchestration, are now somebody else’s job.

That somebody being OpenAI. Developers pay for tokens, tools, and container time; there’s no additional API fee.

The service is organized around four primitives (Agent, Environment, Session, Events), streams progress from durable sessions, and connects to customer-supplied MCP servers. A reference example spins up gpt-6-astra with multi_agent enabled and max_concurrent_subagents: 3, invoking an observability tool over HTTP. Teams can run agents inside an OpenAI-hosted sandbox, self-host via codex exec-server over WebSocket, or pick from nine partner sandboxes at launch: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. OpenAI positions the API as “low” integration effort against “medium” for a self-deployed Agents SDK and “high” for wiring Responses API directly.

Launch customers supplied the marketing numbers, none independently verified: Ciridae reports its evaluation score climbing from 0.71 to 0.85 with a 4x latency reduction on subagent flows; SafetyKit claims 60% lower cost per case after migrating a case-review workflow; Hypha cites 86% fewer failed responses once the harness was separated from the sandbox; Nash.ai says it’s running thousands of long-running agents across global logistics networks on the same infrastructure.

The trade publication byteiota captured the pitch bluntly: “If you’ve ever spent two weeks writing context summarization logic before you could start on your actual agent — this is aimed squarely at that problem.” The same write-up flags what the announcement soft-pedals: US-only data residency, no committed timeline to lift it, and Zero Data Retention unsupported regardless of sandbox. An OpenAI community poster added the operator’s warning: “If you are new to using hosted sandboxes, please be careful when spinning up new containers and calculate the costs first. I learned this the hard way.”

Read this alongside X opening its ads API to third-party AI agents, Wix’s Symphony multi-agent platform aimed at SMB owners, and 6sense piping live buying-intent signals directly into agents, and the shape of the next two quarters gets legible. The distribution pipes and intent signals arrived first. Now the harness arrives. The bottleneck moves from plumbing to product judgment, which is a much harder bottleneck to hide behind.

OpenAI Opens Its Agents API to All Developers, Putting the Codex Harness Behind One Call

Sources