OpenAI’s new Agents API puts long-running agents in the cloud

Taking a task beyond the chat window

Investigating a software outage takes more than a single answer. An assistant needs to examine changes, read errors and test explanations. It also needs somewhere to keep its findings as the work progresses.[2]

OpenAI’s Agents API, launched in public beta on September 10, packages the system behind Codex for developers building that kind of application. It provides the machinery for agents to work with files, run code and coordinate longer tasks.[1]

OpenAI’s examples include an assistant that investigates alerts and requests approval before recovery actions, alongside document reviewers and data analysts. The API lets developers connect those capabilities to their own applications.[2]

Keeping the work moving

Long assignments can outgrow the information a model can hold at once. The service condenses earlier work so an agent can continue, and can divide independent jobs among additional agents. OpenAI calls this supporting software a harness and maintains it alongside its models.[1]

An app can show progress, steer work in flight and continue a session later. Developers choose the tools and execution environment, using OpenAI’s hosting, their own infrastructure or a supported partner.[2]

That division gives an app builder less background machinery to maintain. It also leaves room for products to differ in the information they can access, the jobs they support and how they present the result.

What developers pay for

OpenAI announced no extra fee for the Agents API itself. Model usage, tools and OpenAI-hosted computing environments remain billable, so the cost depends on what the agent uses during a task.[1][2]

The underlying Codex harness is open source. Developers can inspect its code while using OpenAI to operate and maintain the managed service, which launched as a public beta.[1]