Canary

Teams using AI coding agents like Claude Code, Cursor, Codex and Copilot now receive more pull requests than human reviewers can actually read, so review has become a skim. The agents also write their own tests, which pass, but nobody on the team reads or trusts those tests. As a result, breakages surface after merge, as a customer Slack message or a 2am on-call page, rather than being caught beforehand. The core problem the site describes is that the blast radius of an AI-written change is an unknown unknown until it fails in production.

Canary is installed as a CLI (npm i -g @runcanary/cli) and onboarded into a repo via a canary skills command, then triggered either inside the coding agent before a task is marked done, or automatically on every pull request. It reads the diff and the codebase to map every user flow the change can reach, boots the app in a clean sandbox, runs tests designed to break the change, and reports the blind spots as a case file with root cause and a suggested fix. For in-agent runs it snapshots the uncommitted working tree and tests it with nothing committed yet. It integrates with GitHub, Linear, Sentry, Datadog, Notion and Slack, and files results back into the coding agent, PR, or via MCP.

The page frames the problem as new: AI agents now raise more PRs than teams can review, and the same agents write tests nobody reads, making human review and self-written tests insufficient safety nets.