OpenAI introduces the Agents API for long-running cloud agents
OpenAI has introduced the Agents API in public beta, giving developers access to the harness and infrastructure that power its Codex agent. The announcement, published on September 10, 2026, is aimed at teams that want to build agents able to use tools, work with files, coordinate subtasks and keep running across long sessions. OpenAI describes the API as a way to create a cloud agent with one API call by specifying the task, model, tools and execution environment.
The API separates the agent harness from the place where the work runs. Developers can use an OpenAI-hosted sandbox, their own infrastructure or a sandbox partner. OpenAI lists Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel as ecosystem providers. The company says these options can cover different needs for networking, files, secrets, compute, cold starts and cost. OpenAI is also introducing hosted sandboxes managed by OpenAI, with configurable files, packages, skills and plugins.
The Codex connection is the main technical story. OpenAI says the harness manages context, tool use and subagents, and that it will evolve alongside new model releases. The announced capabilities include automatic context compaction for long sessions, tool search to load only relevant tool definitions, programmatic tool calling for parallel or chained operations, MCP and custom functions, and multi-agent support. These features could reduce the amount of orchestration code developers need to maintain, but they do not remove the need to design permissions, retries, observability and human review.
OpenAI says the Agents API is available to all developers in public beta. There are no additional fees for the API itself; customers pay for the tokens and tools their agents use. The company presents customer examples reporting faster workflows, lower costs or fewer failed responses, but these are testimonials rather than independent evaluations. The API is therefore best understood as a new platform layer, not proof that every agent workload will be reliable or economical.
For makers, the release lowers the barrier to building agents that can act over hours or days instead of replying once. For businesses, the choice of environment may make deployment and governance more practical across different security requirements. The important questions will be how much control teams retain over data and execution, how clearly agent actions can be audited, and how performance holds up outside the examples OpenAI selected. Developers can begin with the Agents API overview and quickstart linked from the announcement.