OpenAI Releases Agents API in Public Beta, Offering a Managed Codex Agent Runtime

A developer can now hand off much of the machinery that keeps an AI agent running across a complex task: its session state, orchestration, context management and recovery. OpenAI has released the Agents API in public beta, giving all developers managed access to the Codex-based harness behind that work while they supply the tools and select where the agent executes.

The release changes what an application needs to build itself. Rather than only calling a model and assembling its own agent loop around it, an application can create a durable session, give it a task, receive output events and continue or steer the same session later. OpenAI says developers can follow progress through streamed output or webhooks, including when an agent finishes or needs more input.

A managed loop, with a developer-chosen workspace

The central division of labor is straightforward: OpenAI manages sessions, orchestration, automatic context compaction and recovery. The application provides the agent’s tools and chooses the execution environment. That environment can be an OpenAI-hosted sandbox, a developer-managed environment, an integrated partner sandbox, or no sandbox at all.

A sandbox gives the agent a place to act rather than simply respond. OpenAI says agents operating in one can run code, edit files, connect to Model Context Protocol, or MCP, servers, and produce artifacts. The API supports MCP connections, custom functions and built-in tools such as web search.

Capabilities aimed at longer work

The managed harness is designed for tasks that involve more than a single model answer. Its features include summarizing earlier work to manage the model’s context window, breaking work into subtasks, delegating those subtasks to other agents, and resuming a session where it left off. Those functions address recurring engineering work in agents that must keep track of progress while invoking tools over multiple turns.

• Automatic context compaction can summarize prior work as a session approaches its context limit.
• Programmatic tool calling supports running and combining tool operations in code.
• Multi-agent support can delegate independent tasks to subagents and combine their output.
• The API can connect agents to external data through tools or MCP servers.

OpenAI lists examples that reflect that broader workflow: an incident-response agent that investigates alerts and seeks approval for recovery actions; a data analyst that queries a warehouse with read-only SQL; and a GitHub issue investigator that reproduces reported bugs and shares findings. They are examples of intended applications, not performance evidence for the API itself.

The constraint sits outside the sandbox choice

The launch comes with a material data-handling limit. The Agents API supports data residency only in the United States and does not support Zero Data Retention. OpenAI says choosing a self-hosted sandbox does not make the API eligible for Zero Data Retention. For organizations whose requirements depend on those controls, moving the execution environment outside OpenAI does not resolve the limitation.

There is no separate charge for the Agents API harness itself, according to OpenAI. Usage is billed through the selected model’s API rates, while OpenAI tools and OpenAI-hosted sandboxes carry their standard rates. That makes the product decision less about a new platform fee than about whether to adopt OpenAI’s managed runtime for the workloads a team wants to automate.

OpenAI has also published customer-supplied results, including a Ciridae evaluation-score increase from 0.71 to 0.85 and a fourfold latency reduction in subagent flows, plus SafetyKit’s reported 60% lower cost per case. Those figures are not independent benchmarks, but they point to the claim the new service is asking developers to evaluate: whether a managed harness can reduce the operational work of building agents without giving up too much control.