OpenAI Introduces the Agents API for Long-Running Workflows

OpenAI has introduced the Agents API in public beta, positioning it as a way for developers to build cloud agents around a managed harness rather than assembling every orchestration component themselves. The company says the API exposes the same harness and infrastructure that run Codex, so a single API call that names a task, a model, a set of tools and an environment produces a production-ready agent.

The company says the API lets developers choose a model, tools, and execution environment while supporting longer-running sessions and multiple subagents. Compute can run in an OpenAI-hosted sandbox, on a developer's own infrastructure, or with sandbox partners that OpenAI lists as Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. Tools can be MCP servers, custom functions or built-in tools such as web search, and a multi_agent setting lets a primary agent delegate work to a configurable number of concurrent subagents that each keep their own context.

Two harness features address the hardest part of long-running agents: keeping them useful over hours of work. OpenAI says the API automatically compacts earlier context as a session nears its limit, and a tool-search capability loads tool definitions only when they are needed to save tokens. The company says the harness is maintained alongside each model release and is built on its open-source Codex harness, so developers can read the orchestration logic.

Early customers named in the announcement include Ciridae, Long Lake, WithCoverage, SafetyKit, Dwelly, Hypha, deepsense.ai and Nash.ai. Ciridae CTO Jack Weissenberger is quoted saying that evaluation scores rose from 0.71 to 0.85 and that latency fell to a quarter after adopting the built-in subagent flow.

The practical question is how much operational complexity the managed layer removes in real deployments. OpenAI says the service is available in public beta and that developers pay for the tokens and tools they use. Read the official announcement.