OpenAI Puts Its Codex Agent Harness Behind a Public API

OpenAI announced the Agents API in public beta on September 10, 2026, giving developers direct access to the same harness and infrastructure that powers Codex and enterprise ChatGPT.

What the API Handles Automatically

OpenAI handles session orchestration, context compaction, and recovery — infrastructure work that developers building agent systems have historically had to implement themselves. The API supports long-running tasks, parallel tool calls, and multi-agent collaboration, with agents continuing work across turns while progress streams back to the calling application as it runs.

Built on Open-Source Foundations

The Agents API is built on the open-source Codex execution framework and supports tools such as MCP, custom functions, and web search — giving developers a standardized foundation rather than a fully proprietary black box.

Flexible Deployment Environments

Developers build cloud-based agents by defining a task, model, tools, and compute environment, choosing where the agent actually operates: OpenAI-hosted sandboxes, their own infrastructure, or third-party sandbox partners — a deployment flexibility that lets organizations balance convenience against data residency or infrastructure control requirements.

A Broad Compute Partner Ecosystem

OpenAI formed partnerships with Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel to offer compute environment integrations — a wide roster spanning both established cloud infrastructure providers and specialized sandbox/execution environment startups.

What This Means for Developers Building Agent Systems

Putting a production-grade agent harness behind a single API call — handling orchestration, context management, and multi-agent coordination that developers previously had to build themselves — is a specific signal that the infrastructure layer for AI agents is commoditizing quickly, letting developers focus on their specific agent's task logic rather than reimplementing common agent infrastructure patterns from scratch.