OpenAI opened a public-beta Agents API on September 10, 2026, that exposes the Codex harness as a managed service. Developers create a session with a model, tools, and an environment. OpenAI then keeps the agent running, compacting context, calling tools, and recovering from failures so the work can last hours or days instead of a single chat turn.

The company says useful agents need more than a prompt. They need a harness that manages context, uses tools efficiently, and coordinates subagents, plus infrastructure that can keep a job alive while it writes files, runs code, and stores intermediate results. The API is built around four objects: an agent, an optional environment, a durable session, and the events that flow through a turn. Developer docs say model usage is billed at ordinary API rates, with standard charges for OpenAI tools and hosted sandboxes.

Teams can run the agent in an OpenAI-hosted sandbox, the same class of environment behind Codex and ChatGPT, or on their own machines. OpenAI listed first-class sandbox partners including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. Agents can call MCP servers, custom functions, and built-in tools such as web search, and they can fan work out to concurrent subagents that keep their own context.

Design partners reported concrete lifts. Ciridae said evaluation scores rose from 0.71 to 0.85 and that subagent flows cut latency by 4x. SafetyKit said a migrated case-review workflow cut cost per case by 60 percent. Hypha said separating the harness from the sandbox cut failed agent responses by 86 percent. Nash, which runs long-lived logistics agents, said the API now supplies the session and recovery layer while Nash keeps the tools that touch the physical world.

“The Agents API has enabled us to think differently about how we can architect complex, multi-step workflows. We used to write prompt chains and manage our own set of tool calls, but now we can use agents directly in our code much like how Codex works on your laptop.” Cole Striler, director of engineering, WithCoverage

Decoded Take

This is OpenAI productizing the boring half of agents: sessions, compaction, sandboxes, and subagent orchestration. That is the layer every shop has been rewriting since function calling landed, and it is also the layer that locks a customer to GPT-6 Astra if the harness only shines on OpenAI models. The customer quotes are strong on latency and failure rates, which is the right proof for a beta. The watch item is whether those gains survive once the agent is wired to a customer’s own VPC and MCP servers, not an OpenAI sandbox, and whether a versioned harness actually ships with the next model instead of forcing another rewrite.