The promise of an “agentic browser” is simple: instead of opening twelve tabs, you describe an outcome and the browser does the clicking. The reality, after a week of trying to make six of them earn their keep, is narrower — and more useful — than the demos suggest.

The five tasks

To keep this fair, every browser got the same jobs, on a fresh profile:

  1. Find a train ticket for a specific date and time, and stop before payment.
  2. Compare three products across two retailers and return a table with prices and delivery dates.
  3. Fill a 14-field insurance quote form using a provided JSON file.
  4. Summarise a 42-page PDF that was already open in a tab, with page citations.
  5. Cancel a subscription that hides the button behind two settings pages.

What worked

Comet was the most reliable at task 5 — following links, scrolling, and finding the actual cancel button rather than a “contact us” page. It was also the only one that handled the insurance form without inventing a value for a field it could not find.

Edge Copilot Mode won task 4 comfortably. Citations pointed at real pages in the PDF, and it refused to summarise charts it could not read instead of making something up.

Neither is magic. Both needed a clear instruction with an explicit stopping condition, and both stalled once on a cookie banner.

What did not

The failure mode was always the same: confident progress in the wrong direction. One browser “compared” three products by reading only the first retailer, then produced a plausible table with numbers it had inferred. Another filled a date of birth field with today’s date and carried on.

Task Completed unaided Needed a nudge Wrong but confident
Train ticket 1 / 6 4 / 6 1 / 6
Price comparison 2 / 6 2 / 6 2 / 6
Form fill 3 / 6 2 / 6 1 / 6
PDF summary 6 / 6 0 / 6 0 / 6
Cancel subscription 2 / 6 3 / 6 1 / 6

Task 4 is the reason to care

PDF summarisation being 6 out of 6 is the quiet headline. Reading a long document used to mean either skimming badly or paying for a separate tool. Doing it in the tab you already have, with citations, is the kind of improvement that survives the hype cycle.

Everything else is still a demo with a nice interface — useful when it works, expensive when it does not, and never trustworthy without a human glance at the result.

How I would actually use this today

  • Use it for reading. Long PDFs, dense policy pages, multi-tab research. Keep the citations.
  • Use it for filling boring forms with data you already have in a file, and check every field.
  • Do not use it for anything irreversible. No payments, no cancellations, no “send” buttons.
  • Keep a normal browser profile for banking. Agents that can read the page can read your session.

The category is real, but the interesting part is not autonomy — it is that a browser can now understand what is on the page. That alone is worth the upgrade.