Scope the surface you need
A browser agent operates within browser-accessible interfaces. A computer-use agent can describe a wider interface scope, including desktop or mobile environments when its model, controller, and runtime support them. Product naming is inconsistent, so inspect documented coverage instead of assuming that a broad name implies broad control.
Anthropic now documents a separate browser-use tool path for webpage tasks alongside its desktop computer-use documentation. Google documents multiple environment types for computer use. These examples reinforce the need to examine the exact tool and environment combination rather than extrapolate from a vendor brand.
A browser boundary can simplify the job
For a portal export, the required actions may be limited to navigation, a date filter, and a download. A browser-scoped runtime can be an appropriate starting point. Define what happens to the file after download, because browser control alone does not establish a storage or processing workflow.
For a desktop process that opens a local spreadsheet and a native application, the browser is only one component. You need a supported desktop environment, file access rules, window management, and a way to identify the active application. A browser benchmark cannot validate that whole path.
Prefer explicit handoffs between surfaces
An editorial design for a mixed workflow is to separate collection from transformation. Retrieve the source artifact in a constrained browser session, validate it, and send it to a deterministic file-processing job. Use native GUI control only when the local application offers no suitable supported automation path.
If a desktop agent is necessary, reduce the workspace to the applications and files required for the task. Do not make a personal workstation the default test environment. The important question is what the running process can access, not only what the prompt tells it to ignore.
Evaluate the whole stack
Shortlist by environment, then check the controller, session persistence, deployment, model integration, and outcome evidence. A hosted browser provider can complement an agent framework, while a desktop controller may need an entirely different host. Record unknown capabilities explicitly and ask for a reproducible demonstration before treating them as supported.
A pilot should include a task that crosses the intended boundary: downloading and validating an artifact for a browser workflow, or opening and checking it in the named native application for a desktop workflow. Avoid buying desktop scope for a browser-only requirement unless there is a clear operational reason. Scope should follow the task, not the breadth of the marketing label.