An AI-assisted browser helps you browse; an agentic browser can take the browser work off your hands.

An AI-assisted browser makes the person faster at browsing.

An AI-assisted browser typically sits beside the page and helps with the cognitive part of the task. It can summarize an article, explain an unfamiliar dashboard, translate a passage, answer a question about the current page, or produce a first draft.

Those are useful capabilities. The person still does the operational sequence: open the next tab, find the relevant record, copy the number, update the sheet, and submit the result. The product assists the browsing rather than carrying it out.

An agentic browser owns the operational sequence.

An agentic browser takes a goal with an output: research 30 accounts, reconcile a list of invoices, prepare follow-ups after calls, or complete a supplier comparison. It then opens and reads the necessary pages, extracts information, moves between tabs, and returns the result where it belongs.

The result should be inspectable. A useful run gives you source URLs, a completed table, staged record updates, or drafts ready for review. It should not leave you with a paragraph explaining the manual work that remains.

A four-question test cuts through product labels.

First, who clicks: does the person operate every page or can the product do so?

Second, whose session is used: a clean remote browser or the account context already open on the user’s device? Third, what arrives at the end: advice, or a completed work product?

Fourth, what happens when an action has consequences? A serious agentic workflow should distinguish research and drafting from sending, purchasing, deleting, submitting, or changing records. The label matters less than these operating facts.

Autonomy is useful only when it remains governable.

The value of an agentic browser is not a theatrical cursor.

It is taking repeatable work across the sites and tools where it actually happens. Skills capture a successful process so it can be repeated; routines put it on a schedule when there is a regular trigger.

In Strawberry, the repeatable part can run as a routine while consequential actions remain approval-gated. That model lets a Monday pipeline pass or a competitor scan do the preparation without quietly turning a scheduled check into an unreviewed external change.

Frequently asked questions