On September 10, OpenAI put the Agents API into public beta. The pitch: whatever has been driving Codex internally is now a service anyone can call.
One call instead of a framework
You pass the task, the model, the tools and the environment, and you get an agent that can run. OpenAI handles the rest:
- Context. As a session approaches its context limit, the earlier part gets compacted automatically. That’s the work you otherwise write yourself in a homegrown agent loop, usually too late.
- Tools. Tool search cuts the token cost of a large tool inventory. Programmatic tool calls can run in parallel.
- Subagents. The agent can split a task and hand the pieces to subagents working in parallel.
- Environment. Either OpenAI-managed sandboxes or self-hosted ones via partners like Modal, Cloudflare, E2B and Oracle.
You pay for tokens and tools. No fee for the API itself.
The most telling word in the announcement is a different one, though: OpenAI calls this an evolving Codex harness. Meaning the harness is supposed to get better with every model update without you touching anything. If you’ve ever built your own agent layer on top of a model and then discovered the next model has different habits, you know why that sounds appealing.
The numbers OpenAI cites
From customer deployments: 60 percent lower cost per case, a quarter of the latency, 86 percent fewer failed agent responses. Those are vendor figures about selected customers, not independent measurements. What’s interesting isn’t the size but the selection — cost, latency, failure rate. Exactly the three places where homegrown agent loops fall over.
Why this matters if you use Claude
Anthropic is working the same ground, and has been for a while. Managed Agents shipped in the summer, ant apply added agents-as-code a week ago, and on September 10 the auto permission policy landed. Both companies are now selling the same thing: not the model, but the machinery around it.
The difference is where they started. Anthropic builds its agent platform around permissions and traceability — who may call which tool, who approves, what lands in the log. OpenAI builds around the harness that makes Codex fast. One comes out of the compliance corner, the other out of the performance corner.
I wouldn’t bet on which wins. But if you’re weighing whether to keep maintaining your own agent loop: from here on you’re maintaining it against two vendors who give that work away.
Sources: Introducing the Agents API, Claude Platform release notes