OpenAI launches the Agents API in public beta

OpenAI released the Agents API in public beta on September 10, 2026, giving developers hosted infrastructure to build and run production agents without assembling the harness themselves. The API packages the context management, tool coordination, and long-running reliability that OpenAI says it learned from scaling Codex and ChatGPT for Work to millions of users.

What's new

The Agents API lets a developer spin up an agent with a single API call by specifying a task, a model, a set of tools, and an execution environment. From there it handles the mechanics that normally have to be hand-built:

• Context management that automatically compacts conversation and tool-output history so long-running sessions do not blow past context limits.
• Tool search and programmatic tool calling, so an agent with a large tool library can find and invoke the right tool without every definition being stuffed into the prompt.
• Multi-agent coordination, allowing a primary agent to spin up and manage parallel subagents on pieces of a larger task.
• A choice of compute environments — an OpenAI-managed sandbox, a self-hosted environment, or an ecosystem partner's infrastructure — for agents that need to read and write files or run code.
• Support for MCP, custom functions, and OpenAI's built-in tools, including web search, inside the same harness.

OpenAI says there are no additional fees for using the Agents API beyond standard token and tool-usage costs — it is priced as a wrapper around existing API consumption, not a separate line item.

Context

OpenAI frames the release as the productization of infrastructure it already runs internally. "Useful agents need a powerful harness that manages context, uses tools efficiently, and coordinates subagents," the company wrote, adding that such agents "also need infrastructure that keeps them running reliably for days, with environments where they can work with files, run code, and save intermediate results." That is a direct description of what Codex and ChatGPT for Work already do at scale — the Agents API is that same harness exposed to any developer building on the OpenAI platform, rather than something bespoke to OpenAI's own products.

The release lands the same day OpenAI's developer changelog also confirms GPT-Live-1 has moved to general availability, and comes weeks after OpenAI joined Amazon, Microsoft, Cursor, and Vercel in launching Agent Plugins, an open standard for extending agents with third-party tools. Together the moves show OpenAI building out both ends of the agent stack: an open extension standard for how agents plug into tools, and now a hosted runtime for how agents are actually built and operated.

Why it matters

The market for "agent harness" tooling — the plumbing that keeps a long-running agent on task, within budget, and able to use tools reliably — has so far been dominated by third-party frameworks and orchestration layers built outside any single model provider. By hosting that harness directly inside the API, OpenAI is betting developers would rather not stitch together context compaction, tool routing, and multi-agent coordination themselves, and that keeping it in-house increases the switching cost of building an agent on a competing model. Whether that bet pays off will depend on whether the hosted version is flexible enough to match what custom-built agent frameworks already do for teams with more demanding orchestration needs.