OpenAI Launches Agents API in Public Beta
The API manages long-running agent sessions across OpenAI-hosted or third-party sandboxes with no extra platform fees beyond token and tool usage.
- Sandboxes can run in user VPCs, OpenAI-hosted environments, or providers including Cloudflare, DigitalOcean, E2B, Modal, Oracle Cloud, Vercel, and Daytona
- Built-in multi-agent orchestration splits tasks across parallel subagents that each maintain separate context windows
- Automatic context compaction compresses earlier session data as context limits approach to enable multi-hour runs
- Agents can execute code, manipulate files, install packages, and produce artifacts using configurable CPU, GPU, memory, and storage
- Integrations support the Model Context Protocol, tool search, and programmatic tool calling to filter and chain operations in code
Developers building autonomous agents can offload execution sandboxes, context compaction, and multi-agent orchestration to OpenAI without managing custom runtime infrastructure.
