Draft poster · the finished cut will appear here
What happened
OpenAI announced the Agents API in public beta as the Codex harness and infrastructure exposed through an API. OpenAI says a developer specifies the task, model, tools, and environment while the service hosts and maintains the surrounding run lifecycle.
How it works
The API is built around state rather than a single request. OpenAI says sessions retain durable state and automatically compact earlier context as a run approaches its limit. The announcement lists MCP, custom functions, web search, tool search, programmatic tool calling, and multi-agent delegation as supported paths. Agents can run in OpenAI-hosted sandboxes, on developer-controlled infrastructure, or in supported partner environments.
That packaging addresses the parts of an agent that usually sit outside a model call: how a session resumes, how context is reduced, where tools execute, and what happens when work spans multiple steps. OpenAI says the public beta has no additional Agents API fee beyond the tokens and tools used.
Why it matters
Agent platforms are becoming managed runtimes. Model choice still matters, but durable state, tool plumbing, recovery, context lifecycle, and execution boundaries determine whether a capable agent can do reliable work outside a chat window.
What we don’t know yet
Public beta is not production readiness. OpenAI’s reliability, pricing, and supported-environment descriptions are vendor claims. The announcement does not establish how sessions recover across every failure mode, what each partner environment supports, or how usage costs behave at scale.
Source: OpenAI, OpenAI, September 10, 2026 - original announcement