OpenAI launches Agents API in public beta, exposing the Codex agent harness via one API call

On 2026-09-10 OpenAI released the Agents API in public beta: a managed agent service built on the open-source Codex harness that launches durable, resumable cloud agents through a single API call specifying task, model, tools, and environment. Agents run in…

OpenAI introduced the Agents API in public beta on 2026-09-10 (Reports dated 2026-09-10 from OpenAI News, Hacker News, and MarkTechPost; The Decoder followed on 2026-09-11), giving developers the same agent harness and infrastructure that powers Codex — which The Decoder also describes as the infrastructure behind ChatGPT — through a single API call specifying task, model, tools, and environment. The service is built on the open-source Codex code, providing visibility into model calls, tool use, and context logic, and OpenAI manages orchestration, context compaction, and recovery. Agents can run for hours, execute code, edit files, process files, and connect to MCP servers in OpenAI-managed sandboxes, customer/self-hosted infrastructure, or partner environments from Cloudflare, Modal, E2B, Vercel, Oracle, DigitalOcean, Blaxel, Daytona, and Runloop (The Decoder lists only Cloudflare, Vercel, and Oracle — a subset, not a contradiction). Sessions are durable, resumable mid-task, steerable mid-turn, and can stream output and use webhooks; automatic context compaction lets agents run across hours-long sessions and multiple context windows. Supported tooling includes tool search, programmatic tool calling (parallel calls with results filtered in code) to cut token cost, web search, custom functions, and concurrent subagents that parallelize complex tasks. Documentation examples use the model gpt-6-astra, and suggested applications include an incident response agent, Slack bot, data analyst, and GitHub issue investigator. Billing follows model, tool, and container rates with no additional API fee during the beta; note a minor discrepancy: The Decoder states billing is based solely on token usage, while the other three sources cite model/tool/container rates. Documented limitations: data residency is currently limited to the United States and Zero Data Retention is not supported, even with self-hosted sandboxes. MarkTechPost additionally cites vendor-reported results — SafetyKit cutting case review cost 60% and Ciridae achieving 4x lower subagent latency — and notes scalability learnings drawn from scaling Codex and ChatGPT for Work.

• Launch: Agents API entered public beta on 2026-09-10 per OpenAI News, Hacker News, and MarkTechPost; The Decoder's report appeared 2026-09-11.
• Interface: one API call specifies task, model, tools, and environment; OpenAI manages orchestration, context compaction, and recovery on the Codex harness.
• Harness: built on open-source Codex code, giving visibility into model calls, tools, and context logic; The Decoder frames it as the infrastructure behind Codex and ChatGPT.
• Sandbox options: OpenAI-hosted, customer/self-hosted infrastructure, or partner environments from Cloudflare, Modal, E2B, Vercel, Oracle, DigitalOcean, Blaxel, Daytona, and Runloop; The Decoder lists only Cloudflare, Vercel, and Oracle.
• Agent capabilities: execute code, edit files, process files, and connect to MCP servers; supports web search, custom functions, tool search, and concurrent subagents.
• Programmatic tool calling runs calls in parallel and filters results in code, reducing token cost; context compaction enables hours-long sessions across multiple context windows.
• Sessions: durable, resumable mid-task, steerable mid-turn, with streaming output and webhooks.
• Examples in documentation use the model gpt-6-astra; suggested applications include an incident response agent, Slack bot, data analyst, and GitHub issue investigator.