OpenAI opened the Agents API in public beta on September 10, making the orchestration engine behind its Codex coding agent available to any developer through a single managed API rather than a bespoke, in-house harness.
What the Agents API does
Building a reliable agent typically means writing and maintaining a “harness” — the plumbing that manages multi-turn sessions, keeps context from overflowing as a task runs long, recovers from errors, calls tools, and coordinates subagents working in parallel. The Agents API exposes the same harness OpenAI built to run Codex, so developers supply only the tools and pick an execution environment while OpenAI manages sessions, orchestration, context compaction and recovery.
Agents built on the API run inside a sandbox where they can execute code, edit files, connect to MCP servers and produce artifacts, using either OpenAI’s own hosted sandboxes or outside infrastructure. OpenAI said partners including Cloudflare, DigitalOcean, Oracle, E2B, Daytona, Modal, Runloop and Vercel already offer first-class sandbox integrations. The API also adds native multi-agent support, letting a top-level agent delegate pieces of a task to subagents that run concurrently rather than serializing every step.
Pricing and availability
The Agents API is available today to all developers in public beta, with no separate API fee: usage is billed at the underlying model’s standard token rates, plus standard rates for any OpenAI tools or hosted-sandbox compute an agent consumes. OpenAI said it will keep iterating on the API through the beta period based on developer feedback ahead of a general-availability release.
Why it matters
The launch turns what had been an internal advantage — the engineering behind Codex’s ability to work through long, multi-step coding tasks with limited supervision — into infrastructure any developer can build on. That puts pressure on the growing field of third-party agent-orchestration frameworks and harnesses, since teams that previously had to assemble their own session management, context handling and sandboxing can now rent OpenAI’s instead. It also deepens OpenAI’s push to be judged less on chat quality alone and more on whether its agents can be trusted to complete real, extended work end-to-end — the same competitive terrain where Cognition’s Devin, GitHub Copilot and Cursor are also racing to prove themselves.