OpenAI has made its Agents API available as a public beta. The interface gives developers access to the same underlying infrastructure that powers Codex and ChatGPT, and it is aimed at teams that want to build their own autonomous agents rather than rely on ready-made assistants. The company says there are no additional fees beyond the usual token usage, which keeps the cost model familiar for anyone already working with its models.
What happened
According to the announcement, the Agents API allows developers to create cloud agents that can run autonomously for hours. Such an agent is not limited to a single question and answer: it can work through a longer task, execute code as part of that work, and hand off parts of the job to sub-agents. That structure mirrors how complex projects are usually split between specialists, with one coordinator passing narrower tasks to other workers.
OpenAI is not positioning the service as a closed environment. Cloudflare, Vercel and Oracle are named as providers of additional sandbox environments, so developers can choose where the code produced by an agent actually runs. Sandboxes matter because an agent that writes and executes code needs an isolated place to do it, away from production systems and customer data.
The pricing detail is also notable. OpenAI states that using the Agents API does not add extra fees on top of token consumption. For teams that already pay per token, this removes one more barrier to experimenting with agent-based workflows, since the cost of a pilot can be estimated from the same metrics they already track.
What it means for business
For companies, the arrival of a standard agent infrastructure from a major model provider lowers the technical entry point. Instead of assembling orchestration, code execution and task delegation from separate tools, development teams can start from a single interface and add sandboxing from the vendors that already support it. This shortens the path from an idea to a working prototype, especially where a process involves several steps and several systems at once.
At the same time, longer autonomous runs raise the bar for control. An agent that works for hours and executes code needs clear limits, logging and a human checkpoint before it touches anything critical. Businesses that adopt such tools will benefit most if they first define which tasks can be delegated, what data the agent may access, and who reviews the result. The infrastructure is now easier to reach, but the discipline around it remains the deciding factor.
