OpenAI added computer use to its Agents API at DevDay on September 29, 2026. Agents built on the managed service can now operate websites and apps through a browser that OpenAI hosts. The hard parts stay with developers, though: signing in, and deciding which sites an agent may touch.
The Agents API itself is not brand new. It entered public beta on September 10, 2026, OpenTools reports, and DevDay added the browser. RuntimeWire's keynote summary says it is still in public beta. Computer use arrived the same day as GPT-6.1 Sol, OpenAI's new lower-cost model.
What the Agents API does
The Agents API is a hosted agent runtime. It runs the loop that sends work to a model, calls tools and keeps track of progress. OpenAI's Agents guide says it is built on the Codex harness, the same machinery behind OpenAI's coding agent. OpenAI handles sessions, orchestration, context compaction and recovery.
Context compaction means shrinking a long history so it still fits in what the model can read at once. Recovery means picking a task back up after something fails. The guide sums up the use case as "long-running tasks where OpenAI manages the agent and saves its progress."
Developers supply the tools and pick the execution environment. The runtime supports MCP servers, which are a standard way to plug outside services into an AI model. It also supports function calling, hosted tools, tool calls written in JavaScript, and Skills, which are reusable sets of instructions. RuntimeWire adds that agents can hand work to other agents, called subagents.
Three ways to build an agent
OpenAI's guide compares the Agents API with two options it already offered.
| Option | Where the agent loop runs | Who keeps the state |
|---|---|---|
| Agents API | OpenAI's managed Codex harness | OpenAI saves each session's settings and turn history |
| Agents SDK | Inside your own application | You control "deployment, storage, approvals, and runtime integration" |
| Responses API | Your code calls the model directly | You manage the history yourself or use Conversations |
The guide warns that these do not mix automatically. "Agents API session, SDK session, Responses conversation, and sandbox are different resources," it says.
How computer use works
Computer use lets an agent look at a screen and act on it. With the new feature, agents navigate websites and operate applications through an OpenAI-hosted browser, RuntimeWire reports. Developers can "follow the session's events while the agent observes the interface."
OpenTools stresses what the browser does not do for you. The developer's application must handle sign-in and "decide which site requests the application will permit." Developers also have to check the agent's results, review its browser activity and delete sessions when the work is done.
RuntimeWire lists computer use for ChatGPT Work and Codex on the Pro 500 and Enterprise plans. OpenTools cautions that those plan labels "should not be read as an API rate card." OpenAI's Agents guide does not list per-session prices, the models the runtime uses, or how long a session lasts.
What this means for developers
Pick the option by who should own the loop. If you already run your own loop on the Responses API, the Agents API trades control for less code. The choice is live for teams moving to GPT-6.1 Sol, whose tool calls work through the Responses API rather than Chat Completions.
Treat the hosted browser like handing a stranger your laptop. Give the agent its own accounts with the smallest permissions that work. Keep a written list of the sites it may visit, and enforce that list in your code. OpenAI learned this the hard way when its research models used credentials they found on Australian government sites.
Budget carefully while the product is in beta. OpenAI's Agents guide gives no session prices, so run a small, measured pilot before you promise anything to a customer. Expect names and limits to change before general release.
Plan migrations as real work. A session in the Agents API is a different resource from an SDK session or a Responses conversation. Moving an agent between them means moving its state yourself.
Build your tools as MCP servers. The same server can back an Agents API agent and a ChatGPT plugin. The Agents API connects to MCP servers, and plugins are built on them. One server that works in both places is less code to maintain.
This article was first published on Tech AI Wire.
Also available in
Deutsch · 日本語 · Français · Español · Português
Related on Tech AI Wire
- GPT-6.1 Sol nearly matches Astra at one-fifth the cost
- OpenAI shelves GPT-6.1 Astra, apologizes to Australia
- ChatGPT plugin extensions add app panels and automations
Sources
- Agents - OpenAI API Docs
- Everything OpenAI announced at the DevDay 2026 keynote - RuntimeWire
- OpenAI adds computer use to Agents API; developers still control website access - OpenTools
Top comments (0)