OpenAI has launched the public beta of its Agents API, providing developers with a managed cloud environment for AI agents. The service leverages the execution framework and infrastructure originally developed for Codex, allowing for the creation of applications capable of handling complex, multi-step tasks.
The API enables agents to execute code, process files, and manage intermediate results over extended periods, ranging from hours to days. Developers can utilize OpenAI-hosted sandboxes or deploy agents on their own infrastructure. Partnerships have been established with providers such as Cloudflare, DigitalOcean, Modal, and Oracle to offer various hardware configurations.
Key features include advanced context management, which automatically compresses information to maintain continuity across multiple context windows. The system also supports programmatic tool calling and parallel execution, allowing agents to load tool definitions on demand and coordinate tasks efficiently.
The platform supports multi-agent architectures, enabling developers to break down complex projects into sub-tasks handled by concurrent agents. OpenAI provides configurations for up to three parallel sub-agents. The underlying execution framework is open-source, allowing developers to review the core logic.
The Agents API is currently available to all developers in public beta. OpenAI does not charge extra fees for the API itself, with costs limited to standard model token and tool usage. The company plans to iterate on the service based on developer feedback before a full release.
