OpenAI announced the Agents API in public beta on September 10, 2026, presenting a managed agent harness derived from Codex. According to OpenAI, the service provides an API for building long-running, production-ready cloud agents that can work with files, run code, and preserve state across extended sessions.
Capabilities and deployment options
OpenAI said developers can create an agent with a single API call by specifying task, model, tools and environment. The company highlighted the example use of model “gpt-6-astra” and described support for tools such as MCP, custom functions and built-in tools like web search. The API also accepts configurations such as “multi_agent” with parameters like “max_concurrent_subagents”: 3 in the example provided by OpenAI.
Deployment choices include OpenAI-hosted sandboxes, self-hosted infrastructure, or partner sandboxes. OpenAI named partners including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. The hosted sandboxes are described as managed environments where agents can access files, packages, skills and plugins.
Execution model and tooling
OpenAI said the Agents API uses a versioned Codex harness that the company maintains and evolves alongside model releases. Features called out include automatic context compaction to keep agents working across long sessions, tool search to load relevant tool definitions on demand, and programmatic tool calling to run operations in parallel or chain them in code.
The API supports multi-agent workflows by delegating subtasks to subagents that maintain their own contexts while the main agent coordinates results. OpenAI presented the approach as a way to parallelize research, analysis and coding tasks without requiring developers to build custom orchestration.
In customer feedback cited by OpenAI, Jack Weissenberger, CTO of Ciridae, reported an evaluation score rise from 0.71 to 0.85 and a 4x latency reduction after adopting the Agents API’s subagent flows.
OpenAI said Agents API is available in public beta to all developers. During the beta there are no additional fees for the Agents API itself; users pay for tokens and tools their agents consume. OpenAI indicated it will iterate on the service based on developer feedback during the beta period.
Original source: OpenAI News