Introducing the Agents API

OpenAI has announced the launch of the Agents API in public beta, providing developers with a managed service designed to build and run cloud agents. According to OpenAI, the service is powered by the Codex harness, which provides infrastructure for orchestration, long-running sessions, and tool use. Developers can create a production-ready agent in a single API call by specifying the task, model, tools, and environment. OpenAI hosts and maintains the harness, while developers can choose their compute environment from an OpenAI-managed sandbox, their own infrastructure, or various ecosystem partners such as Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. The Agents API introduces features like automated context compaction to handle sessions spanning hours, tool search to reduce token usage and cost, programmatic tool calling for parallel operations, and multi-agent support that delegates tasks to subagents. The open-source Codex harness allows developers to inspect core logic. There are no additional fees for using the Agents API during the public beta, with developers paying only for the tokens and tools used.