OpenAI Agents API Launches in Public Beta for All Developers
Developers can now build long-lived cloud agents using the same infrastructure behind OpenAI's own products, without additional API fees. Current limitations including US-only data residency and no Zero Data Retention support restrict use in regulated workloads.
OpenAI released the Agents API in public beta on September 10, 2026, giving all developers access to the same agent harness and infrastructure that powers Codex and ChatGPT through a single API. The API manages the agent loop on OpenAI's infrastructure, handling model calls, tool use, and context. Developers can run agents in an OpenAI-managed sandbox, their own infrastructure, or one of nine partner sandboxes from providers including Cloudflare, DigitalOcean, Vercel, Modal, and others. Capabilities include automatic context compaction, tool search for reduced token usage, programmatic parallel tool calling, multi-agent coordination, secure code execution, file modification, and MCP connectivity. The SDK supports both Python and TypeScript, though its small surface area means developers must implement advanced behaviors themselves. OpenAI stated plans to iterate rapidly on the API and expand availability over time.
Coverage on September 14 confirmed MCP connectivity and secure code execution as Agents API capabilities omitted from the launch summary, and documented Python and TypeScript SDK support.
A Hacker News discussion found developers already building LLM-agnostic alternatives to the API, with commenters arguing that independence from any single lab is a competitive advantage over OpenAI's tightly coupled offering. OpenAI stated plans to iterate rapidly on the API and expand availability over time.
OpenAI released the Agents API in public beta on September 10, 2026, opening access to all developers.
The API, built on the open-source Codex harness, handles context management, tool use, and subagent coordination through a single interface, with agents runnable in an OpenAI-managed sandbox, on developer infrastructure, or across nine partner sandboxes including Cloudflare and Vercel. Pricing carries no flat fee: costs combine per-turn model token charges, per-call fees for hosted tools such as web search, and container time for Code Interpreter. Current limitations include US-only data residency and no Zero Data Retention support, restricting regulated workloads.
MIT researcher Beatriz Yankelevich's use of GPT-5.6 Sol and Codex for quantum computing experiments published
Wired reported unanswered questions about the Hugging Face incident, including why early warnings were not escalated and why monitoring failed.
OpenAI published a technical report on its investigation into the Hugging Face incident involving its agents.
I send one email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free