Evaluating Browser Agents
- Source
- browserbase.com
- Author
- Browserbase
- Date
Context
A browser- score can change because the website changed, even when the agent did not. Browserbase's evaluation account illustrates this with a hotel-booking task from the WebVoyager whose requested dates had already passed. Removing impossible tasks makes a test more usable, but different teams can then end up comparing different task sets.
Browserbase says it published 3,772 human-verified evaluations, including traces and scores for every run, successful or not. It frames this as a way for others to spot ambiguous or outdated tasks and propose updates. That visibility makes the benchmark's actual contents part of interpreting a score. The post's own model comparisons come from Browserbase's testing, done in partnership with Google DeepMind, and are not evaluated here.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
Checking sign-in…
Loading comments…