OpenAI opens Agents API public beta, putting the Codex harness behind a managed call

OpenAI opened a public beta of the Agents API on September 10, 2026, giving applications access to the Codex harness through an OpenAI-managed API. OpenAI runs the agent loop on its own infrastructure, coordinating model calls, tool use and context, while the calling application supplies the tools an agent can reach and decides where its work actually executes. A Hacker News submission titled "AI coding — OpenAI Agents API" linking the API overview drew 50 points and 44 comments.

The product sits between two interfaces OpenAI already documents. With the Responses API, the application owns the loop: it receives function calls, runs them, returns the output and calls the model again. With the Agents SDK, an in-process runner performs the tool loop, switches agents after handoffs, and stops when a run finishes or pauses for approval, with sessions, tracing and guardrails attached. The Agents API takes that loop off the developer's server and adds what OpenAI's docs list as automatic context compaction, multi-agent orchestration, programmatic tool calling and support for MCP servers.

The managed piece shipped alongside the beta is the sandbox. OpenAI's documentation says agents can execute code, edit files, connect to MCP servers and produce artifacts inside an execution environment. In the hosted option, OpenAI provisions and connects a Linux workspace stocked with Python, Node.js and command-line tools, while the application supplies the task and retrieves the results. Developers who need their own image, their own compute or a private network can point the harness at a self-hosted sandbox.

OpenAI describes the split as harness against compute. The harness is the control plane around the model and owns the agent loop, model calls, tool routing, handoffs, approvals, tracing, recovery and run state. The sandbox is where files change, commands run, ports open and provider state lives. OpenAI says the service covers long-running sessions, parallel tool calls and multi-agent collaboration, and that later runs can reconnect to earlier work through saved run state and session snapshots rather than rebuilding a workspace.

Pricing keeps the parts separate. The developer-community announcement says there are no additional fees for using the Agents API, and that customers pay for the tokens and tools their agents consume. OpenAI-hosted sandboxes are billed at standard container rates, and model usage is billed separately at the selected model's API rates. One early commenter on the announcement thread asked specifically for a sandbox pricing page and, after finding it, advised developers to calculate container costs before spinning up environments.

Beta status carries its own limits. The sandbox documentation states that API details, defaults and supported capabilities may change while the feature is in beta, which means the execution environment and the orchestration layer are both moving targets for anyone building against them now. OpenAI frames the self-hosted path as the answer for teams that need control over the sandbox image, the underlying compute or network isolation, and the hosted path as the default for teams that want the envialed for teams that want the environment handled for them.

What remains unstated is the timetable. OpenAI describes the Agents API as available in public beta to all developers, with no date given for general availability, and the guide offers a quickstart that creates an agent which writes and runs a script in an OpenAI-hosted sandbox. Whether developers move orchestration onto OpenAI's infrastructure, or keep the loop and adopt only the sandbox, is the choice the beta leaves open.