Bug Manager shows how agents can turn a pile of defects into one ranked list a team can burn down. Agents scan every repository, classify what they find, rank it, log a ticket for each finding, and attempt a fix by opening a pull request. A team then spends a sprint reviewing from the top of the list.
Everything here is simulated. The repositories, findings, pull requests, and tickets are generated in your browser. No code is scanned, nothing is sent anywhere, and no language model is called. The real system ran Claude agents over real repositories, and its findings are confidential.
Ranking is illustrative. A finding’s risk score is its severity weight × (1 + how exposed clients are to the repository) × the agent’s confidence. Open any finding to see its score worked out.
Not every finding gets a pull request. When the agent’s confidence is low, or the fix needs a design decision, the finding is flagged for an engineer. That rule is illustrative too.