OpenAI turns the Codex harness into a public Agents API

OpenAI has launched the Agents API in public beta, turning the harness behind Codex into a managed developer product. The company says the API lets developers create long-running agents by specifying a task, model, tools and environment while OpenAI manages sessions, orchestration, context compaction and recovery.

The release is aimed at teams that want production agents without building the surrounding runtime themselves. Agents can run in OpenAI-hosted sandboxes, in customer infrastructure, or through partner sandbox providers including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. OpenAI says hosted sandboxes can work with files, run code, use packages, load skills and produce artifacts.

The API also packages several mechanics that have become central to modern coding agents. It supports automatic context compaction for long sessions, tool search to load relevant tool definitions only when needed, programmatic parallel tool calls, MCP and custom functions, plus multi-agent support for delegating work to subagents.

For builders, the important shift is that OpenAI is no longer only selling models and separate tool primitives. It is offering the agent runtime itself: durable sessions, execution environments and orchestration patterns that previously had to be assembled in-house. The beta has no separate platform fee beyond model, tool and hosted sandbox usage.