OpenAI is renting out the machinery behind its own coding agent. The Agents API, in public beta since September 10, is the harness under Codex. It’s the part that keeps an agent working for hours, manages what it remembers, finds its tools and splits work among helpers, and now it’s a service anyone can call.
You create an agent in one API call: the task, the model, the tools, and where it runs. The harness does the unglamorous work. It compacts older context as a session fills, so the agent keeps going across several windows. It loads tool definitions only when they’re needed. It can hand pieces of a job to subagents that run in parallel, each with its own context, without you writing the orchestration. Tools can be MCP servers, your own functions or built-ins like web search. The agent runs in an OpenAI-hosted sandbox, on your own infrastructure, or with one of the partners OpenAI lists: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. The harness itself is the open-source Codex harness, so you can read the logic.
The pricing line is the one that stands out: “you simply pay for the tokens and tools your agents use.” It’s a beta, and OpenAI says it’ll change quickly on the way to general availability. What a day-long agent actually costs in tokens isn’t on the page.