On September 10 (local time), OpenAI announced the public beta launch of the Agents API, allowing developers to access cloud-based AI agent execution environments managed by OpenAI via API calls.

Leveraging the agent execution framework and infrastructure originally developed for Codex, this service enables developers to specify tasks, models, tools, and execution environments to build AI applications capable of performing multi-step tasks.

OpenAI states that developers can create production-grade agents with a single API call. These agents can execute code, process files, and save intermediate results in the cloud, while continuously performing tasks that span hours or even days. OpenAI handles the maintenance of the agent execution framework, allowing developers to focus on the specific tools, knowledge, and workflows required for their applications.

The Agents API supports OpenAI-hosted sandbox environments as well as deployment on a developer's own infrastructure or within partner-provided sandboxes. OpenAI has partnered with providers such as Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel to offer a variety of CPU, GPU, memory, and storage configurations.

Additionally, the "OpenAI Hosted Sandbox" service has been introduced, enabling developers to configure files, software packages, skills, and plugins for their agents without the need to build their own sandbox infrastructure. This environment—which utilizes the same sandbox infrastructure as Codex and ChatGPT—is designed for executing code, processing files, and generating outputs.

To handle long-running tasks, the Agents API offers context management capabilities; it automatically compresses earlier information as a session approaches context limits, allowing agents to continue tasks across multiple context windows. The API also supports tool discovery and programmatic tool invocation, enabling agents to load tool definitions on demand, execute calls in parallel, chain operations, or filter results to minimize unnecessary context usage.

Multi-agent support is another key feature of the Agents API. Developers can break down complex tasks into independent sub-tasks processed in parallel by sub-agents, with a primary agent coordinating and aggregating the results. OpenAI's examples demonstrate configurations allowing for up to three concurrent sub-agents.

The Agents API is built upon the open-source Codex execution framework. OpenAI notes that developers can view the public codebase to understand the core logic governing model calls, tools, and context management, while OpenAI remains responsible for maintaining the framework and continuously enhancing agent capabilities alongside model updates. The Agents API is now available for public beta to all developers. There are no additional API usage fees; users pay only for the model tokens and tools they utilize. OpenAI has stated that it will continuously iterate based on developer feedback during the public beta and gradually move toward an official release.