The new product is infrastructure, not a new kind of model

OpenAI has introduced an Agents API in public beta. Its pitch is not simply that developers can call a model. They could already do that. The new service is intended to host the machinery around a longer-running agent: a session that persists, tools that are coordinated, an environment for files and code, and a record of what happened along the way.

OpenAI says this is the same broad harness work it has developed for Codex. The company describes useful agents as needing more than a capable model. They need context management, reliable tool use, a place to work and a way to keep going when a task lasts longer than one response.

That makes the launch a practical story. Agent demos often hide the unglamorous part: starting a process, storing intermediate work, recovering from a failed tool call and keeping permissions intelligible. Those details determine whether a promising prototype can be used by a team.

What OpenAI says developers get

According to OpenAI's announcement and API documentation, a developer can create a session by specifying a task, model, tools and an environment. The platform can provide a hosted sandbox, while developers can also use their own environment. The documentation lists sessions, files, MCP connections, web search and functions among the pieces that can be composed around an agent.

OpenAI says the API is available to all developers during the public beta. It also says there is no separate charge for the Agents API itself, although users pay for the tokens and tools an agent consumes. That distinction matters. A managed control plane may be free to enter while a long-running workflow still generates meaningful model, search, sandbox or third-party costs.

The company has published the Codex harness as open source. That gives developers something concrete to inspect. It does not make every hosted feature interchangeable, and it does not answer every question a security or platform team will have about logging, retention, deployment location or external integrations.

The real design question is still authority

Hosted infrastructure can make it easier to give an agent more time and more tools. It should not make it easier to give the agent broad authority by accident. An agent that can read a project, browse internal systems and execute code needs a clear boundary around each of those powers.

The useful pattern is usually narrower than a grand autonomous assistant. Give the system a defined job, limited credentials and a review point before it changes money, customer data, production settings or anything else difficult to undo. Record what it used and why. Test the workflow against ordinary failures as well as dramatic ones.

Those are engineering and governance choices, not features that arrive automatically with a managed API. The launch can reduce the work of operating an agent platform. It cannot outsource accountability for what an organisation lets the agent do.

A public beta is an invitation to test the seams

OpenAI says it will iterate during the beta. That is worth taking literally. Long-running agents create awkward cases: an expired credential halfway through a job, a tool result that changes the plan, a user who revises the request, or a task whose costs rise faster than its value.

For developers, the sensible first question is not whether an API can produce an impressive demonstration. It is whether the session model, observability, security controls and pricing fit a real workflow with a clear owner. The answer will differ between a bounded research task and an agent that can act across a company's systems.

OpenAI's release adds a significant managed option to a crowded agent-tooling market. The practical test now is whether teams can make these systems easier to supervise, not merely easier to start.

What is confirmed, what OpenAI says, and what remains open

Confirmed: OpenAI published the Agents API announcement and accompanying developer documentation on 10 September 2026. The documentation describes sessions, environments, tools and an OpenAI-hosted sandbox option. OpenAI lists the API as public beta and says access is available to all developers.

OpenAI's claims: the managed API brings the harness and infrastructure behind Codex to developers, can support reliable multi-day work, and makes production agents easier to build. Those are product claims from the company, not independent measurements of reliability, security or cost in every use case.

Open questions: how the service performs under sustained production load, how teams will configure least-privilege access in varied environments, which safeguards will prove easiest to audit, how costs behave in long tasks, and how much operational portability developers will have across hosted and self-hosted setups.

Sources

  1. OpenAI — Introducing the Agents APIPrimary product announcement, 10 September 2026. Source for the public beta, the Codex-harness framing, available access and OpenAI's description of managed agent infrastructure.
  2. OpenAI API documentation — Agents API overviewPrimary technical documentation. Source for Agents API concepts including sessions, environments, hosted sandboxes, tools, MCP connections and observability.
  3. OpenAI API documentation — PricingPrimary pricing reference used to distinguish the stated lack of an extra Agents API fee from token and tool charges.