Codex's execution layer becomes a developer API
OpenAI introduced the Agents API in public beta on September 10. Developers specify a task, model, tools and environment, while OpenAI operates the harness used behind Codex and ChatGPT Work. Work can run in an OpenAI-hosted sandbox, a team's own infrastructure or an integrated environment from partners including Cloudflare.
The release is primarily an operations layer rather than another model endpoint. It compacts earlier context as sessions approach their limits, loads only relevant tool definitions through tool search, and supports MCP, functions and web search. Programmatic tool calling can filter and combine results in code, while multi-agent support delegates independent work to subagents. OpenAI says the API itself has no surcharge; users pay for model tokens, tools and compute they consume.
A managed harness does not inherit business accountability
The service is a public beta, so interfaces and behavior can change. Cost and latency improvements in the announcement are early customer reports, not independent guarantees across workloads. A sandbox and harness also do not decide which production permissions are appropriate or verify that an irreversible action is correct.
Teams should begin with read-only access, require human confirmation and idempotency for payments, messages and deletion, and test budget ceilings, cancellation and recovery for long sessions. Managed infrastructure should not be confused with permission for unattended execution.