A browser agent uses the web to finish a task, not just explain one.

A browser agent turns a browser task into an outcome.

A browser agent receives an outcome, works through the required pages, and returns the finished or prepared result. For a supplier-price review, that can mean visiting ten supplier portals, recording each current price, comparing it with last month’s sheet, and flagging changes worth checking.

The browser is important because many business processes do not have one clean API. They live across dashboards, directories, forms, portals, and internal tools. A browser agent is designed to carry the work across that uneven surface.

Page understanding is more useful than blind clicking.

Older browser automation follows a fixed script: click this coordinate, then type into that field. It is fast when a page never changes and fragile when a button moves or a form adds one question.

A browser agent instead has to identify what it is looking at, decide which item matters, and verify whether the last step worked. That does not make it infallible. It does make it suitable for web work where the page structure and the exception cases change from run to run.

Good instructions name the evidence and the stopping point.

“Find competitors” is a research request, not yet an executable job.

“Find 25 UK payroll platforms with a self-serve pricing page, record the starting price and URL, and flag any claim you cannot verify” gives the agent a target, evidence standard, output shape, and a reason to stop.

That is also why a reusable skill helps. Once a sourcing pass has a working brief, source rules, columns, and QA checks, saving those instructions means the next run starts from a proven operating definition rather than a blank chat.

Oversight belongs at the point where the work changes something.

Reading public pages and compiling a private draft are usually low-risk.

Sending a message, changing a customer record, submitting a form, making a purchase, or deleting information are different: the action has a real external effect.

A sensible browser-agent design separates the research and preparation from the decision to execute. In Strawberry, the companion can do the legwork in the browser and hold consequential steps for approval. A recurring version can run as a routine, but the guardrail stays attached to the action rather than disappearing because the task is scheduled.

Frequently asked questions