Make Better
List your asset
AI agents & workflow automation

OpenAI Agents API gets computer use: what changes for browser agents

•Make Better Editorial

OpenAI?s Agents API can now run tasks in a hosted browser. Here?s what changes, where approvals matter, and when it fits better than deterministic automation.

OpenAI added computer use to the Agents API on September 29, giving developers a managed way to let agents navigate websites and interact with browser interfaces. The browser runs in an OpenAI-hosted environment while the application starts the session, supplies the task, follows events and handles access decisions.

What changed

The Agents API provides a managed agent harness with tool calling, tool search, multi-agent work and context compaction. Computer use adds a browser execution layer for UI-driven tasks such as testing a website, collecting information or using an application when a direct API integration is not the right interface.

Make Better analysis

Browser automation and agent reasoning can now live in the same managed session. That reduces infrastructure work for teams building agents, but browser execution is still less deterministic than a stable API workflow. Use it where visual navigation or changing interfaces are part of the problem, not as a default replacement for reliable integrations.

How the hosted browser loop works

  1. Create a browser session and keep its session ID.
  2. Send the task and follow session events while the agent works.
  3. Handle website-origin access requests before the browser visits a new origin.
  4. Handle authentication in the application when an account is needed.
  5. Verify the result and review browser activity before ending the session.
Origin approval is not per-action confirmation

OpenAI says the hosted browser requires approval before accessing each new website origin. Approving an origin does not guarantee confirmation before every consequential action. Applications that need stricter controls should restrict what the hosted browser can reach or use a browser runtime they control.

Browser agent or deterministic workflow?

Choose by task shape

Use computer use when?Prefer APIs or workflows when?
The job depends on navigating a web UIA stable API or webhook exposes the needed action
The interface requires visual interpretationThe process must execute identically every time
A human can review uncertain stepsThe action is high-volume or tightly controlled
The task mixes reasoning with browsingLatency and predictable branching matter more than flexibility

A useful hybrid pattern is to keep stable actions in APIs or workflow tools and give the browser agent only the steps that genuinely require UI navigation. That limits the surface area where page changes or ambiguous controls can create errors.

A safer rollout pattern

  1. Start with read-only research, QA or data-collection tasks.
  2. Allow only the website origins the task needs.
  3. Keep account access handled by the application.
  4. Require human review around high-impact changes.
  5. Measure successful completion, recovery rate and human correction.
Bottom line

Computer use makes the Agents API practical for workflows that still live behind browser interfaces. The strongest use case is giving an agent a managed browser for steps that cannot be expressed cleanly through APIs, while keeping consequential actions bounded and reviewable.

Sources & useful resources