OpenAI Launches Agents API: Codex Harness as Managed Service

OpenAI launched the Agents API in public beta on September 10, exposing the same Codex harness that powers Codex and ChatGPT for Work as a managed, cloud-native API. A single call can now launch a durable agent session with its own compute environment, file access, tool use, subagent delegation, and automatic context compaction — infrastructure that previously took teams weeks to build.

Key facts

• Release: Public beta launched September 10, 2026; available to all developers immediately.
• Core endpoint: POST https://api.openai.com/v1/agents/sessions (requires OpenAI-Beta: agents=v1 header).
• What it manages: Session state, orchestration, context compaction, crash recovery, tool execution, sandbox I/O, subagent coordination, long-running session control (hours to days).
• Four core objects:
  • Agent — model + instructions + tools + MCP servers
  • Environment — sandbox (OpenAI-hosted or partner/self-hosted)
  • Session — durable, long-running agent instance
  • Events — streaming outputs for progress tracking and steering

What it does for you

The Agents API handles the “boilerplate” that every serious agent project must build:

• Session state & orchestration — durable sessions that survive interruptions, with automatic recovery
• Context compaction — as sessions approach token limits, the harness compacts earlier context independently for root agent and each subagent, no custom summarization needed
• Subagent coordination — spawn parallel subagents with separate working contexts (max_concurrent_subagents configurable)
• Tool & MCP integration — built-in tools (web search, code interpreter, file search), MCP servers, custom functions
• Sandbox I/O — code execution, file read/write, artifact production in isolated environments
• Streaming events — real-time progress tracking and mid-session steering

Execution environments: your sandbox or theirs

Developers choose the execution environment per session:

• OpenAI-hosted sandboxes — fully managed containers, billed at standard container rates
• Partner environments — Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel (9 partners at launch)
• Self-hosted — your own infrastructure, connected via the API; you manage compute, connection, shutdown, and file persistence

Pricing: no platform fee

• No Agents API fee during public beta
• Model usage — billed at the selected model’s standard API rates
• OpenAI tools — standard tool rates (web search, code interpreter, file search, etc.)
• OpenAI-hosted sandboxes — standard container rates
• Partner/self-hosted sandboxes — billed through respective providers

Important constraints (beta)

• Data residency: US-only during beta
• Zero Data Retention (ZDR): not supported in beta, even with self-hosted sandboxes
• Regional availability: US-only data residency for managed sandboxes
• Rate limits: standard model-level limits apply

What this means for builders

The Agents API productizes the “harness” layer that separates toy agents from production systems. Teams no longer need to build context compaction, crash recovery, subagent coordination, or sandbox management — OpenAI runs the harness; you supply the task, model, tools, and environment choice.

Early adopters report concrete gains: Ciridae reports 4x latency reduction, SafetyKit 60% cost reduction, Hypha 86% fewer failed responses. But two hard blockers for regulated/European teams: US-only data residency and no ZDR support during beta (even with self-hosted sandboxes, ZDR is not available).

What to watch

• Whether ZDR support arrives before GA — a blocker for regulated industries
• Whether data residency expands beyond US
• How partner sandbox pricing compares to OpenAI-hosted
• Whether Anthropic or Google launch equivalent managed harness APIs

Official source

OpenAI — Introducing the Agents API (September 10, 2026); OpenAI Developer Community — Introducing the Agents API and hosted sandboxes (September 10, 2026); Agents API Overview.