Intel / repo

browser-use 1.0.5 — browser-use 1.0.5

Source: github.com
Author: sarath-menonTop Viber
Date: Published Nov 14, 2026

Context

Browser Use 1.0.5 adds an automatic check at the end of every run. After the browser agent finishes, the library sends one more low-effort request to the same model that drove the run. That model compares what happened against the original task and fills in `result.judgement`, a structured verdict on whether the run succeeded, why it failed if it did, whether it saw a , and whether a bot block or CAPTCHA got in the way (release notes). The judge is on by default. Developers can pass their own success criteria with `judge_=` or turn it off with `judge=False`. Per the merged pull request, each judged run costs exactly one extra request, and that request's tokens are added to the run's reported usage and cost.

The judge sees the agent's trace, its last screenshots and the final page, unless vision is disabled. It also writes "gotchas": short notes recorded whenever a site rejected input, threw an error, or blocked or rate-limited the agent, even when the run succeeded. The release gives the example "The form requires phone numbers in (###) ###-#### format." That is the kind of detail a developer would otherwise find only by reading a long trace. The maintainers report a small check: 13 real runs, each judged three times. Gotchas appeared in all 15 verdicts where the site pushed back, compared with 6 of 15 before the change. They appeared in 2 of 24 verdicts on clean runs, and success verdicts did not change. That is the project's own small-sample measurement, not an independent evaluation.

The release also brings back the 0.x options `use_judge` and `ground_truth`, now as aliases for `judge` and `judge_context`. That matters because the 1.0 rewrite deleted most of the old library and had the agent write Python directly against the Chrome DevTools Protocol. Code that passed those two options works again, though other 0.x judge interfaces such as `judge_` stay removed. The version-bump pull request calls 1.0.5 the first PyPI release with the judge.

Key takeaways · AI-distilled

browser-use 1.0.5 judges every run: one extra low-effort request to your own model checks the run against the task and fills result.judgement with success, failure reason, prompt-injection and bot-block flags, and gotchas. judge=False turns it off.
The judge reads the trace, the run's last screenshots and the final page, and records gotchas whenever a site rejected input, errored, blocked or rate-limited the agent, including on runs that succeeded.
Across 13 real runs judged three times each, the release notes report gotchas on 15 of 15 verdicts where the site pushed back, up from 6 of 15, and on 2 of 24 clean runs, with success verdicts unchanged.

Terms in this piece · Glossary

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.
prompt injection — An attack that hides instructions in content an AI will read — a webpage, email, or document — tricking it into following the attacker instead of the user.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.

Read the source github.com