With AI agents writing more of the code, PRs pile up, and the reviewer is often the first person to open the changed page and find that it does not work. browser_review_gate makes “it works in the browser” a condition for asking a human to review. A GitHub Action looks at each PR, decides whether it changes what a user sees or does, and if so takes the review request back. The author runs one command in their AI coding agent (Claude Code, Cursor, Codex): the agent opens the app in a real browser with Playwright, goes through the scenarios, and posts a report to the PR with the pages, what it saw, screenshots and traces. Once the report covers the change, the same reviewers are asked again automatically. The human still reviews the code, but no longer spends that time discovering the feature is broken. The gem checks reports strictly and can verify each trace against the tested commit, so a skipped run does not slip through. It works with or without an AI key in CI, and one command sets a Rails app up.