OpenAI launches Agents API for building long-running cloud agents

OpenAI has launched the Agents API, a managed service for building and running AI agents that keep working in the cloud rather than on a single device or browser session.

The service targets long-running cloud agents: programs that can keep executing a task after the person who started them has closed the app or logged off. OpenAI says the API handles orchestration, the coordination of the steps and tools an agent uses to finish a job, along with tool use and stateful sessions, meaning an agent retains context across a task instead of starting over each time it is called. Tool use, in this context, means an agent can call outside programs, such as running code or searching for information, to help complete what it has been asked to do.

According to OpenAI, the Agents API is powered by what the company calls the Codex harness, the underlying system that provides orchestration for the new service. OpenAI’s announcement does not explain how the harness works internally, whether it is the same system used elsewhere at the company, or what separates it from other approaches to running agents.

What the announcement leaves open

OpenAI’s post, titled “Introducing the Agents API,” is brief. It does not state a price for the service, a rollout timeline beyond the announcement itself, or which developers get access first. It does not name any company or product already built on the API, and it does not describe usage limits, supported programming languages, or what a developer needs to integrate an existing application with it. The announcement also does not say whether the Agents API works alongside OpenAI’s other developer tools or replaces an earlier way of running agents that the company may have offered before. Anyone deciding whether to build on the service will need OpenAI’s fuller documentation, which the announcement does not link to directly.

The launch reflects a wider pattern among AI companies of packaging agents as a hosted cloud service instead of software that developers must run and maintain on their own servers. A managed service moves the work of keeping an agent’s infrastructure running, the servers and processes needed to keep a long session alive, from the developer’s side to OpenAI’s.

What to watch next: whether OpenAI follows up with pricing, access details, and technical documentation for the Agents API, and whether developers who get early access publish concrete examples of agents built on the Codex harness, since OpenAI’s announcement names none.