OpenAI said on Sept. 10 U.S. time that its Agents API has entered public beta, giving all developers access to the agent harness and infrastructure behind Codex and ChatGPT Work through a single API.
The company said developers can create production-ready cloud agents by specifying a task, model, tools, and execution environment in one API call. OpenAI hosts and maintains the harness. It was the third developer product introduced by OpenAI on the same day, following the GPT-Live-1 voice API and a financial services edition of ChatGPT.
Three execution environment options
OpenAI said different workloads require different compute, storage, and deployment choices, and the Agents API lets developers choose the sandbox setup that fits their needs.
The first option is an OpenAI-hosted sandbox. It uses the same sandbox infrastructure already used by Codex and ChatGPT, is configured and managed by OpenAI, and can load a developer’s files, packages, skills, and plugins.
The second option is deployment on the developer’s own infrastructure.
The third option is an ecosystem partner environment. The initial group includes Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. These options offer fully managed deployments or deployments inside a VPC, along with specific file and secret storage mechanisms and different CPU, GPU, and memory configurations.
Automatic context compression and parallel sub-agents
OpenAI said each new model release often forces developers to rewrite their harness, taking time away from improving the application itself. Agents API provides versioned access to these capabilities, and the harness is maintained alongside model updates.
OpenAI highlighted three recently added capabilities.
- Context management: when a work session approaches the context limit, earlier content is automatically compressed. Developers do not need to build their own compression logic, and workflows can span multiple context windows.
- Tool search and programmable tool calling: relevant tool definitions are loaded only when needed to reduce token usage. Programmable tool calls also let agents run tasks in parallel, chain related actions, and filter or merge results in code. Supported tools include MCP, custom functions, and built-in web search.
- Multi-agent support: complex tasks can be split into separate parts and handled by parallel sub-agents, each with its own context, while a main agent coordinates and combines the results.
OpenAI said the Agents API is powered by the open-source Codex harness. Developers can inspect the core logic for coordinating model calls, tools, and context in the public code repository, while OpenAI handles operations and maintenance.
No additional fee during the beta
The Agents API is available in public beta starting now for all developers. OpenAI said it will not charge an additional fee. Developers will pay only for the tokens and tools used by the agent, based on the pricing page, while sandbox container usage will be billed separately by the environment provider.
OpenAI also said it plans to iterate quickly during the beta based on developer feedback, with the goal of moving toward general availability. The company asked developers to report what is working well, where friction remains, and what else is needed to build and run agents in production.
Follows OpenAI’s push toward autonomous task completion
Chain News previously reported that GPT-6 Astra, released in early September, emphasized autonomous task completion. With Agents API, the harness and sandbox required to run those longer tasks are now offered as a service, so developers no longer need to assemble orchestration frameworks on their own.