MEDIA PREVIEW
Overview & result
Fastbrowse indexes live page controls, lets Jev choose the next browser action, uses a generative LLM for planning and reading, and attaches verbatim page quotes to answers. Its repository includes an animated demo and detailed author-run comparison receipts.
From the original post
A browser agent that picks instead of generating.
How Jev fits in the loop
- Observe the page and enumerate valid action candidates.
- Ask Jev to pick an action from those controls; use a generative model where reading or planning is needed.
- Execute in Chrome, verify outcomes in code and cite page text for answer claims.
How to reproduce
- Review the public evaluation protocol and per-run receipts in `docs/evals.md`.
- Run fastbrowse with a Jev key, planning-model key and local Chrome or a Browser Use Cloud browser.
- Repeat the same tasks against the comparison agent and separate action selection from planning cost and browser time.
Why this build matters
An evidence-rich separation of Jev decision, LLM generation and deterministic browser execution, with explicit failure reporting.
Limitations
- The README calls the project pre-alpha and not recommended for production.
- Its reported 41/42 pass rate and cost figures are author-run on a specific 14-task suite, not a general benchmark.