OpenAI Launches Agents API Public Beta With Managed Codex Harness

OpenAI introduced the Agents API in public beta on September 10, 2026, packaging the harness and infrastructure behind Codex as a managed developer service. The company announcement and platform docs state that OpenAI runs session orchestration, long-running context management, and recovery while applications supply tools, instructions, and an execution environment.

The pitch is a single API call that creates a durable agent session. OpenAI’s launch example configures a model (gpt-6-astra in the sample), an MCP server tool, multi-agent settings with up to three concurrent subagents, an OpenAI-hosted sandbox, and a task prompt. Requests use the OpenAI-Beta: agents=v1 header; current SDKs add it automatically.

Harness and environments

Per the Agents API overview, the managed Codex harness can run commands in a sandbox, apply skills, connect tools or MCP servers, steer an in-flight session, compact prior work as context fills, delegate to subagents, and resume later. Environment choices are openai_hosted, self_hosted (via codex exec-server inside the developer’s compute), or partner sandboxes. The September 10 announcement lists first-class integrations with Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. OpenAI also notes that the core Codex harness remains open source on GitHub.

Pricing and data controls

OpenAI and its developer-community announcement say there is no additional fee for the Agents API itself; billing follows standard model, tool, and hosted-container rates. Documentation is explicit that the Agents API currently supports data residency only in the United States and does not support Zero Data Retention—even when execution is self-hosted. Customer quotes on the launch page cite vendor-reported gains (for example Ciridae on evaluation score and latency, SafetyKit on cost per case); treat those figures as vendor-supplied testimonials, not independent benchmarks.

This brief is about the managed Agents API runtime, not GitHub Copilot’s earlier GPT-6 Astra model availability.

Primary sources are OpenAI’s September 10 “Introducing the Agents API” post, the Agents API overview and quickstart on developers.openai.com, and the OpenAI Developer Community announcement of the API and hosted sandboxes.