OpenAI has added computer-use capabilities to the Agents API, allowing AI agents to interact with websites through an OpenAI-hosted browser. The update moves agent development closer to workflows in which a model can not only reason about a task and call APIs, but also operate software through a user interface.
What OpenAI added
The September 29, 2026 API update adds computer use to the Agents API. An agent can operate inside a hosted browser and interact with web applications. Application developers remain responsible for website access approvals and sign-in flows, which is important because browser automation can perform consequential actions.
Why browser control matters
Many business applications expose APIs, but a large number of useful workflows still live behind graphical interfaces. Computer use provides another route for those systems. Instead of requiring a custom integration for every website, an agent can navigate the interface in a manner closer to a human operator.
This is especially relevant for research, back-office operations, testing, data entry, and workflows involving services that do not provide suitable APIs. It also creates a new class of engineering problems around authentication, permissions, confirmation, and recovery.
How developers should design around it
Applications should keep sensitive authentication and approval logic under application control. Browser actions should be scoped to the minimum permissions required, and important transactions should have explicit confirmation or review steps. Logging is also important so teams can reconstruct what an agent did.
Relationship to agentic development
Computer use is one part of a broader Agents API architecture. Developers can combine model reasoning with tools, hosted execution, and multi-agent patterns. The result is a system that can plan a task, use software, inspect the result, and continue based on what it observes.
Practical takeaway: Teams evaluating computer-use agents should begin with reversible, low-risk workflows and measure task completion, failure recovery, latency, and approval requirements before allowing agents to perform consequential actions.
Source: OpenAI API Changelog