OpenAI Agents API: Codex's harness becomes a managed service

OpenAI launched the Agents API public beta on September 10, 2026, offering developers the managed Codex harness and a choice of execution environments. The announcement opens a new decision for teams building agents: how much of the machinery that keeps an agent working should they operate themselves? OpenAI’s announcement makes the beta available to all developers and introduces OpenAI-hosted sandboxes alongside partner integrations.

The important change is the unit developers hand over. A model request asks for the next response. Here, an application hands a task to a persistent session, with OpenAI managing the ongoing work around model calls. This is a news analysis of the launch and its documentation, checked September 11; it is not a production benchmark.

The product is the machinery around the model

A harness is the software that turns model outputs into continued work: deciding which tools to call, feeding results back, managing context and coordinating subtasks. The Agents API overview assigns sessions, orchestration, compaction and recovery to OpenAI. Developers supply capabilities and choose where code runs.

The launch also describes automatic context compaction, tool search, programmatic tool calling and parallel subagents. These address familiar problems: a long investigation fills its context, a tool catalog grows unwieldy, or several independent checks need to run together. They do not establish that every application becomes faster or cheaper.

One useful distinction appears in the multi-agent documentation: subagents have separate contexts but share the environment’s filesystem. Separate reasoning space is not file isolation. Two agents editing the same file still need coordination.

A sandbox choice is not a choice of who runs the harness

The architecture documentation separates three components:

The sandbox is optional. An agent using remote tools may need no local workspace. When it does, developers can use OpenAI-hosted compute or connect their own environment. A third-party sandbox changes where commands execute; it does not relocate the Agents API harness into that provider’s cloud.

The documented provider list contains nine integrations: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle Cloud Infrastructure, Runloop and Vercel. Teams still need to evaluate the chosen provider’s deployment options and costs. Nine integrations do not imply nine identical security, hardware or billing configurations.

One call starts the work; it does not finish the integration

The quickstart uses client.beta.agents.sessions.create(...) to configure an agent, submit input and stream progress. Raw HTTP requests require OpenAI-Beta: agents=v1. Project keys need agent-session permissions plus inference permission.

For a developer, the attractive part is having a session to continue instead of rebuilding a task loop around every response. But a successful create response is not proof that the workspace is ready. The hosted sandbox guide distinguishes provisioning from a connected environment; failed setup can prevent work from starting.

A realistic first application might review a document, check evidence with tools, and produce a report. Its developer still defines what a correct report looks like, how users see progress and which actions require approval. An API can supply execution machinery without supplying those product decisions.

No extra platform fee does not mean a free agent

The announcement says there is no additional charge for using Agents API. The billing categories are model usage, applicable tools and hosted containers. Reducing that to “tokens and containers” omits tool charges.

The current pricing page lists container rates per 20-minute session, per container:

It also states that eligible container sessions are billed by the minute with a five-minute minimum. That qualification matters: use the applicable billing treatment rather than treating the table as a universal minimum charge. Model and tool usage remain additional; third-party compute follows its provider’s billing.

For budgeting, count the whole completed task: coordinator and subagent tokens, tool use, retries and environment lifetime. A faster parallel run may use more total resources. Compare cost per accepted result before inferring savings from the absence of an API surcharge.

Recovery still has a boundary

The most revealing detail is in the self-hosted sandbox lifecycle. A mid-turn disconnect can fail a tool; reconnecting does not automatically restart a killed command. Reusing an environment ID does not restore files on replacement compute. Deleting a session does not stop external compute.

Those rules narrow the word “managed.” OpenAI owns the session machinery, while the application remains responsible for the external environment’s lifecycle and durable files. The security guidance likewise leaves credential placement and access boundaries with developers. An agent can read credentials exposed inside its environment.

This is where application engineering remains valuable: business permissions, reliable tool handlers, useful evaluation and clear responsibility when something fails.

Open source underneath, hosted operations on top

The Codex repository license is Apache-2.0 licensed. Its open-source history also predates 2026: OpenAI announced the open-source Codex CLI on April 16, 2025. That establishes an earlier public foundation, not that every feature of today’s managed service existed then. The September launch does not substantiate an “August 2026 first open-source release” timeline.

Our analysis is that OpenAI is competing to operate more of the application runtime. An inspectable harness lowers the barrier to understanding its design; a managed service offers an alternative to maintaining it. Model improvements can then arrive alongside changes in how the service uses tools and context.

That makes runtime behavior, session interfaces and operational integration potential sources of switching cost. It does not prove an industry-standard control plane has already emerged, or that independent harnesses have lost their value. Teams needing different model providers or deeper control over execution policy still have reasons to own more of the stack.

For a concrete comparison of locally operated tools, see our coding-agent harness comparison. For this new service, start with the official quickstart and evaluate one bounded workflow. An OpenAI-compatible model endpoint alone does not establish support for the separate Agents API session service.