Hand it to AI Browser

Examples and evidence

Examples, with a clear evidence boundary.

The three workflows above illustrate the intended experience. They are not completed real-world task records. Judge a task by its files, page feedback and remaining work.

These workflows are being validated in the development version. The public download may not include them yet; check the version and service status below.

Collect & download

The goal
Find two public reports on renewable energy from the past year. Download the PDFs and keep their source links.
Along the way
Check the publisher and date, compare versions, and skip duplicates. A login page or an unreadable file is not a finished download.
The result
Two readable PDFs with source links. One older edition excluded; one restricted report remains unavailable.
Explore a task

Forms & surveys

The goal
Use my saved details and preference for quiet activities to fill this weekend survey. Leave it ready for my review.
Along the way
Match the requested fields to saved details, check each page after filling, and ask together about anything missing. The request stops before submission.
The result
Contact details filled and three preferences selected. The optional comment is blank; the survey has not been submitted.
Explore a task

Explore by preference

The goal
Find independent art websites with thoughtful writing and a quiet visual style. Show me a small selection to refine.
Along the way
Visit candidates and compare them with your preferences. Feedback such as “more like this” changes the next search; the agent cannot decide that you are satisfied.
The result
Two candidate links with reasons. A calm editorial site is the strongest fit; a shop-heavy option is excluded. Continue from your feedback.
Explore a task

Earlier measured runs

These development records remain separate from the new workflows and the current download. They do not establish an overall completion rate.

Historical measured runs, listed separately by task scope
Task and versionObserved resultWhole-task conclusion
25-field practice task2026.09.15 · code2225 / 25 blanks passed; 3 errors corrected without human answers.Completed
Continuous playback2026.09.15 · code23The first two parts ended naturally; part 3 played for about 31 seconds; an added title was incomplete.Report incomplete
100 fields across three pages2026.09.14–15 · code20The first two pages passed; 7 blanks remained on page 3 and no completion report was produced.Not completed

These are single runs from different development versions and cannot be combined into an overall success rate.

View test conditions and evidence rules

All three records came from real page execution in the Android browser using qwen3.8-flash. The 25-blank and continuous-content runs received no human webpage-operation help; the long run preserved its incomplete result.

Start with one useful task.

For a first task: find one useful public PDF on a site you know, download it, then check the file and its source.

Start here