OpenAI released the Agents API in public beta on Sept. 10, exposing the harness and infrastructure that runs its Codex coding agent to any developer through a single API, according to OpenAI's documentation and a report from MarkTechPost.
The API is built around four objects: an agent (a model, instructions, tools and MCP server connections), an optional environment sandbox for running code and editing files, a durable session that persists across turns, and the events a session streams back, OpenAI's documentation says. Environments can run on OpenAI's own infrastructure or on one of nine partner platforms, including Cloudflare, Modal, Vercel and DigitalOcean.
OpenAI said the service handles context compaction, tool search to cut token costs, and delegation to independent subagents automatically, so developers do not have to build that orchestration layer themselves. There is no separate API fee; usage is billed at standard token, tool and container rates, according to the documentation.
Early users cited by MarkTechPost reported concrete gains: Ciridae said its evaluation score rose from 0.71 to 0.85 with four times lower latency, SafetyKit reported a 60% cost reduction per case, and Hypha said it saw 86% fewer failed responses.
OpenAI is betting that most teams building agents don't want to hand-roll session state, sandboxing and subagent coordination, and is willing to give that infrastructure away for the price of the tokens it burns. That's a direct challenge to the growing pile of open source agent-harness projects doing the same job, several of which showed up on the developer wire again today.