OpenAI turns the Codex harness into a managed Agents API

OpenAI launched the Agents API in public beta for all developers on September 10, 2026. The API exposes the managed harness and infrastructure used by Codex: developers specify a task, model, tools and execution environment, while OpenAI maintains the orchestration layer. The API supports long-running sessions with automatic context compaction, tool search, programmatic tool calling, MCP and custom functions, built-in tools such as web search, and parallel subagents. Execution can run in an OpenAI-hosted sandbox, on customer-managed infrastructure or through first-class sandbox partners including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. OpenAI says the API itself adds no separate fee beyond the tokens and tools consumed.

Agent builders have increasingly had to own two products at once: the useful workflow and the harness that keeps a model working across tools, context windows and parallel tasks. The Agents API productizes that second layer. Teams can adopt a maintained Codex-style control loop without giving OpenAI exclusive control of execution infrastructure, because the environment remains selectable. That can reduce orchestration work, but it also moves more application behavior onto a versioned vendor harness whose semantics can change during beta. Builders need to separate model behavior, harness behavior and sandbox behavior when testing reliability and cost.

The Agents API does more than expose another model endpoint. OpenAI operates the control layer that manages context, tools and subagents, and developers interact with it through agent sessions. The launch explicitly describes this as the same class of harness and infrastructure used to run Codex.

OpenAI says the API automatically compacts earlier context as a session approaches its context limit, preserving information needed to continue. That lets applications span multiple model context windows without implementing their own compaction loop, although teams should still test what information survives compression on their workloads.

Tool search can load only relevant tool definitions rather than placing every schema into the prompt. Programmatic tool calling can run calls in parallel, chain operations and filter results before returning them to model context. The API supports MCP, custom functions and built-in tools such as web search.

Developers can use OpenAI-hosted sandboxes, run the agent against their own environment or select integrated sandbox providers. That separation matters for networking, secrets, compliance, performance and cost: using OpenAI’s harness does not require every command or file operation to execute inside an OpenAI sandbox.

The API can let one agent delegate work to multiple subagents with bounded concurrency. OpenAI positions this as a built-in way to parallelize larger tasks rather than forcing application code to maintain a separate orchestration graph for every branch of work.

OpenAI says it will maintain and improve the harness alongside model launches and expose versioned access to capabilities. That can reduce application maintenance, but it also means model upgrades and harness upgrades are separate variables. Production teams should pin, evaluate and audit both rather than assuming an API-compatible model change leaves agent behavior unchanged.