OpenAI opens its agents API to developers

OpenAI released the Agents API in public beta, exposing through a single managed API the harness and infrastructure that runs its own Codex coding agent. OpenAI operates the agent loop on its own infrastructure — coordinating model calls, tool use and context — rather than requiring developers to build session orchestration themselves.

The API’s stated features include automatic context compaction, which summarises earlier exchanges as a session approaches its context limit so an agent can continue across multi-hour tasks without an application managing that state, and multi-agent delegation, letting a primary agent hand independent subtasks to subagents that keep their own context while the parent coordinates the results. Developers choose where the agent executes from a list of first-party sandbox integrations including Cloudflare, Modal, Vercel and Oracle Cloud, alongside several smaller providers and OpenAI-hosted sandboxes, with pricing limited to standard token and tool costs rather than a separate API fee.

OpenAI published early case studies alongside the release: the fraud-detection firm SafetyKit reported a 60% lower cost per case after moving its review workflow onto the API, Hypha said separating the agent harness from its sandbox cut failed agent responses by 86%, and Cirridae reported a four-times latency reduction on subagent workflows alongside a rise in its evaluation score. These are OpenAI-selected customer accounts rather than independently verified benchmarks.

The release turns infrastructure OpenAI had kept internal — the scaffolding behind Codex and ChatGPT Enterprise’s agentic features — into a general-purpose product, arriving as agentic coding and computer-use tools proliferate across labs and as OpenAI itself has been reporting internally that its own coding agents now outwork its human researchers on some tasks.