OpenAI released the Agents API in public beta on September 10, 2026. It gives applications access to OpenAI’s managed Codex harness - the agent loop behind its Codex coding agent - through an API. In OpenAI’s words, developers “build agents with a managed Codex harness while OpenAI handles session orchestration, context compaction, and recovery.”
The API is organized around four concepts: an agent (model, instructions, tools and MCP servers), an optional environment (a sandbox or computer where the agent runs code and edits files), a durable session that persists across turns, and the events and items that flow in and out. Built-in features include delegation to subagents, automatic summarization of earlier work, programmatic tool calling, web search and MCP connections. Model usage is billed at standard API rates, and OpenAI tools and sandboxes at their own standard prices. The documentation notes the API currently supports US data residency only and does not offer Zero Data Retention.
This is a shift in what the model vendor sells. Until now, most teams building agents wrote their own harness - the loop that calls the model, runs tools, trims context and recovers from failures - or adopted an open-source framework. OpenAI is now offering that layer as a hosted product, much as model access itself became an API product. It competes directly with agent frameworks and with the orchestration layers other vendors ship.
The launch does not show that a managed harness produces better agents than a custom one, and beta status means behaviour and pricing can change. The lack of Zero Data Retention and non-US residency will also keep some regulated customers away for now.