OpenAI launches Agents API beta for cloud-based multi-agent tasks

OpenAI has launched a beta version of its Agents API, letting developers call cloud-based AI agent environments managed by OpenAI. The service reuses the agent execution framework behind Codex, allowing developers to specify tasks, models, tools, and runtime environments to build AI apps that handle multi-step tasks.

Agents can run code, process files, save intermediate results, and keep working on tasks for hours or even days in the cloud. The API supports OpenAI-hosted sandboxes as well as developer-owned infrastructure or partner sandboxes. OpenAI has partnered with Cloudflare, DigitalOcean, Oracle, Vercel and others to offer different CPU, GPU, memory and storage setups.

The API also offers context management, automatically compressing early information as sessions near context limits. Multi-agent support lets developers split complex tasks into subtasks handled in parallel, with up to 3 concurrent sub-agents. The beta is free to use beyond model token and tool fees.