OpenAI launched the Agents API in public beta on September 10, 2026, providing developers with the managed agent harness and infrastructure used by Codex. The API is available to all developers with no additional usage fee beyond the tokens and tools their agents consume.
The company said developers can create an agent with a single API call specifying its task, model, tools and environment. OpenAI hosts and maintains the harness, while developers select where the agent runs: in an OpenAI-managed sandbox, on their own infrastructure or through a sandbox partner.
OpenAI is introducing its own hosted sandbox using the infrastructure that powers Codex and ChatGPT. According to the company, the service provisions and manages an environment in which agents can run code, work with files and produce artifacts; developers can configure it with their files, packages, skills and plugins. OpenAI also named Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel as integration partners offering options including managed environments, VPC deployments, storage mechanisms and different compute configurations.
Long sessions tools and subagents
The Agents API includes context management for work spanning multiple context windows. OpenAI said the system automatically compacts earlier context as a session approaches its limit while preserving information needed to continue, reducing the need for developers to implement their own compaction logic. The company also said it will provide versioned access to harness capabilities with model launches and continue updating the harness alongside its models.
For tool use, the API can load relevant tool definitions as needed and supports programmatic calls that run in parallel, chain operations, and filter or combine results in code. Supported options include Model Context Protocol servers, custom functions and built-in tools such as web search. Multi-agent support can divide a complex task into independent assignments for parallel subagents, each with its own context, while a main agent coordinates and combines their results.
OpenAI said the service is built on the open-source Codex harness, whose public codebase exposes the core logic for coordinating model calls, tools and context, while OpenAI operates and maintains the hosted implementation. The company plans to iterate during the public beta before general availability. Early customer results presented by OpenAI included Ciridae reporting an evaluation-score increase from 0.71 to 0.85 and a fourfold latency reduction, SafetyKit reporting a 60% reduction in cost per case, and Hypha reporting 86% fewer failed agent responses after separating the harness from the sandbox; these figures are customer claims published in OpenAI’s announcement.