OpenAI on September 10, 2026 opened public beta access to its Agents API, exposing the managed Codex harness — session orchestration, context compaction, tool selection, and multi-agent delegation — behind a single API call. The company is pitching it as the same infrastructure that runs Codex and ChatGPT for Work, now available to any developer for building long-running cloud agents.
Codex Harness As A Service
The Agents API lets developers create a production-ready agent in one API call by specifying the task, model, tools and execution environment. OpenAI hosts and maintains the harness that keeps agents working for hours or days; developers pick a sandbox — OpenAI-hosted, self-hosted inside a VPC, or from partners including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. Pricing is model tokens plus paid tools, with no additional API fee.
Automatic Compaction, Tool Search, Subagents
The harness now automatically compacts earlier context as sessions approach their limits, so agents can carry work across multiple context windows without custom bookkeeping. A new tool search loads relevant tool definitions on demand — cutting token usage while protecting the prompt cache — and programmatic tool calling lets an agent execute tool calls in parallel, chain them, and filter results in code. Multi-agent support decomposes tasks into subagents that run in parallel while the primary agent coordinates.
Codex Foundation, Open Codebase
OpenAI says the Agents API is powered by the open-source Codex harness. Early customers cited in the announcement include Ciridae, Long Lake, WithCoverage, SafetyKit, Dwelly, Hypha, deepsense.ai and Nash.ai, with Ciridae claiming a 4× latency reduction from out-of-the-box subagent support and Hypha reporting an 86% drop in failed responses after moving to the managed harness.
Why It Matters
The pitch is that developers stop re-implementing prompt chains, custom orchestration, and their own recovery logic. It also tightens the loop between OpenAI's model releases and the harness those models expect — capabilities the company has previously kept inside Codex are now versioned and exposed through the API. See our coverage of Anthropic's live-database security test with Claude and Accenture and Google Cloud's Gemini Enterprise business group for context on the broader agentic-AI push, alongside the Know-Your-Agent interoperability standard.
Reporting based on coverage from OpenAI, MarkTechPost, and Blockchain.News.