A browser agent operates the browser. You state the outcome

How is it different from a chatbot or copilot?

Three separations. Actor: a chatbot advises while you operate, and an agent operates while you review. Scope: a sidebar sees the page you are on, and an agent traverses many tabs and sites in one task, including the logged-in ones where your work lives.

Deliverable: a chatbot returns text in a panel, and an agent returns changed state: the filled sheet, the updated record, the staged email. The wider taxonomy is untangled at agentic browser vs AI browser.

What does one do on a real task?

The instruction: "collect this month’s receipts from the inbox and the four vendor portals, and file the expense report." The companion searches the inbox for receipts, visits each vendor portal in your already-signed-in session to download the invoices, and extracts date, amount, VAT and category from each.

It enters the lines into the expense tool in another tab and stops. The submission is consequential, so it waits on an approval card while you glance over the 23 lines. One instruction, one review, one click. A broader inventory of job shapes is in use cases.

What separates a real agent from a demo

Properties that never show in a launch video.

It works in your accounts without taking your passwords. It pauses before irreversible actions and shows the actual payload. It flags what it could not do instead of narrating success. It runs the same task on a schedule without degrading.

And you can watch finished runs of real tasks rather than curated clips. Ours are published at real runs, approval pauses included, and the login model is on the security page.

How to run the test yourself

Take one task you would hand a capable temp, write it as sources, steps and what done looks like, and watch who does the clicking.

If the browser works while you do something else and the finished thing is waiting for your approval, you are looking at an agent. Strawberry is free for Mac and Windows.

Frequently asked questions