OpenAI Opens Up Agents API: Agent Orchestration Out of the Box

OpenAI has opened up its Agents API. A single call sets the task, the model, the toolset, and the environment the agent runs in. Under the hood sits the same orchestration that powers Codex and the agents inside ChatGPT.

What has changed

Until now, the loop connecting a model call, tool calling, memory, and retries was assembled by hand on top of libraries. That layer now moves to the provider: automatic context compaction and session management are both advertised. In practice, a conversation no longer breaks against the context window partway through a long chain of steps, and the platform itself holds the state between steps. The product competes directly with Anthropic's managed agents.

What this means

A team that set aside a quarter for plumbing of its own is back to the build-or-buy question. This share of the development budget gets cut or dropped entirely, and the timeline to a first working version is worth recalculating. The check is quick: take one process that is already broken down into steps and build it on ready-made orchestration — there is a clear benchmark, since the plan for the in-house version is already costed.

What remains unclear

Pricing and limits. Availability and a legal payment route from Russia. Requirements for storing session data. Whether there is an option to run inside a company's own perimeter or to choose a region.