OpenAI has released a public beta of the Agents API, giving developers managed infrastructure for cloud agents that can continue working for hours. The service is based on systems used by Codex and ChatGPT and is intended to remove some of the orchestration work behind long-running tasks.
The API manages context automatically, can invoke tools in parallel and lets one agent delegate work to sub-agents. Built-in capabilities include web search, code execution and file processing, while developers can add custom functions or tools exposed through the Model Context Protocol. The underlying agent harness builds on the open-source Codex project.
Developers can run tasks in sandboxes hosted by OpenAI or use environments from Cloudflare, Vercel and Oracle. OpenAI says there is no separate fee for the orchestration layer; billing is based on the model tokens consumed. That makes the cost sensitive to how long an agent runs and how many sub-tasks it creates.
Public beta status means teams should treat interfaces and behavior as subject to change and apply their own permission, spending and review controls. The release nevertheless turns capabilities previously embedded in OpenAI products into a programmable service, allowing developers to build persistent agents without assembling context management and task delegation from separate components.