OpenAI launches Agents API in public beta, exposing the managed Codex harness for long-running cloud agents

OpenAI has released its Agents API in public beta: a managed service built on the open-source Codex harness that runs durable, resumable cloud agents with code execution, file editing, MCP connectivity, parallel tool calling, and concurrent subagents in…

OpenAI launched its Agents API in public beta, reported on 2026-09-10 by MarkTechPost and covered by The Decoder on 2026-09-11 (Hacker News item the same day described the documentation). The API exposes the same infrastructure that powers Codex and ChatGPT as a managed service built on the open-source Codex harness: OpenAI handles orchestration, context compaction, tool search, and recovery, while agents can run for hours, execute code, edit files, and process files. Sessions are durable and resumable mid-task, can stream output and use webhooks, and are steerable mid-turn. Supported capabilities include programmatic tool calling (runs calls in parallel and filters results in code), web search, custom functions, MCP server connectivity, and concurrent subagents. Agents run in OpenAI-hosted sandboxes, self-hosted environments, or partner sandboxes; MarkTechPost names nine partners (Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel), while The Decoder lists only Cloudflare, Vercel, and Oracle. Documented example applications include an incident response agent, a Slack bot, a data analyst, and a GitHub issue investigator, using the model gpt-6-astra. Vendor-reported results cited by MarkTechPost: SafetyKit cut case review cost by 60% and Ciridae achieved 4x lower subagent latency. Constraints: data residency is currently US-only and Zero Data Retention is unsupported even with self-hosted sandboxes. Sources disagree on billing granularity: Hacker News and MarkTechPost say pricing follows model/token, tool, and container usage with no extra fee, while The Decoder reports no extra fees and billing based solely on token usage.

• OpenAI released the Agents API in public beta (reported 2026-09-10 by MarkTechPost; The Decoder coverage dated 2026-09-11)
• The API is a managed service built on the open-source Codex harness and runs on the same infrastructure as Codex and ChatGPT
• OpenAI manages orchestration, context compaction, tool search, and recovery; agents can run for hours
• Sessions are durable, resumable mid-task, steerable mid-turn, support streamed output and webhooks
• Capabilities include programmatic tool calling (parallel execution with in-code result filtering), web search, custom functions, MCP server connectivity, and concurrent subagents
• Sandbox options: OpenAI-hosted, self-hosted, or partner sandboxes; MarkTechPost lists Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel, while The Decoder names only Cloudflare, Vercel, and Oracle
• Example applications: incident response agent, Slack bot, data analyst, GitHub issue investigator; examples use the model gpt-6-astra
• Vendor-reported results: SafetyKit cut case review cost by 60%; Ciridae achieved 4x lower subagent latency