OpenAI documents an Agents API for durable cloud agents

OpenAI published official documentation for an Agents API that runs durable cloud agents with managed sessions, sandboxes, MCP connections, and subagents.

Long-running agent work is now a documented platform surface — sessions and sandboxes, not a single chat completion.

OpenAI published Agents API documentation for durable cloud agents — long-running workers that keep going after one chat turn — on a managed Codex harness.

The page says OpenAI manages sessions, orchestration, context compaction, and recovery. The app supplies tools and picks an environment.

Agents can run in a sandbox, an isolated workspace, edit files, connect to MCP (Model Context Protocol) servers, and produce artifacts.

Example snippets set model to gpt-6-astra and show MCP, web_search, programmatic_tool_calling, and multi_agent with max_concurrent_subagents: 4 — a cap of four helper agents running at once.

Pricing on the page: model usage, OpenAI tools, and hosted sandboxes at their standard rates.

The API currently supports data residency only in the United States — keeping stored data in U.S. regions — and does not support Zero Data Retention, the option to avoid keeping customer content.

No discrete dateline on the overview hero. We will not invent one.