On 10 September OpenAI launched the Agents API in public beta. With one call you create a cloud agent with a given task, model, tools and environment, and behind it sits the same harness that powers Codex. There is no extra fee: you pay for tokens and tools.
- You choose the environment: an OpenAI sandbox, your own infrastructure or a partner such as Cloudflare, Vercel, Modal, E2B or Daytona.
- The harness compacts earlier context in long sessions, loads tools as needed and runs subagents in parallel.
- The Codex harness is open source; OpenAI maintains and updates it alongside its models.
The harness is the part of an agent nobody talks about on stage. The model is the star. The harness is the stage, the lighting and the people backstage who know who goes on when.
Anyone who has built an agent from scratch knows that is where the weeks go: keeping the context from overflowing, loading tools without burning tokens, splitting work between subagents and collecting the results. OpenAI has already done it for Codex. From 10 September it gives it to everyone.
What you gain and what you give
You gain time. A harness someone else maintains and updates alongside the models means a new model doesn't force you to rework the harness yourself.
You give something quieter: your agent lives in someone else's house. The environment can be yours, the harness code is open and you can read it, but OpenAI sets the pace of change. If a harness improves with every release, it also changes with every release.
The good news is that the choice of environment is real. If your files and code must stay with you, you run the agent on your own infrastructure and use the harness as a service, although the model itself still runs at OpenAI.
If you have already written your own harness and it works, don't rush to throw it away. Build the same agent both ways and measure the cost per task. The number will tell you more than the beta.