6 min read1258 wordseng
The author argues AI can improve code quality by using multiple models (Claude sub-agents, Codex, Cursor Bugbot) to find, rank, and validate bugs—including KISS/DRY violations, inaccessible HTML/JSX, and missing SQL indexes—so humans can fix critical issues first or even abandon
- • The author argues that AI coding tools can be used to write higher-quality code more slowly, not just to produce fast low-quality output.
- • LLMs and agents are presented as especially effective at finding bugs in codebases, including subtle bugs in recent models from Anthropic and OpenAI.
- • The main challenge is prioritizing and validating bug findings, so the author uses multiple models together to reduce hallucinations and false positives.
- • The author describes a workflow where agents find and rank bugs in a PR, then humans verify the findings and write a final report.
- • In practice, the author fixes critical and high-priority issues first, skips lower-value fixes when appropriate, or abandons a PR if it reveals a flawed approach.
- • This slower, more methodical use of AI may not increase raw productivity, but the author says it improves code quality, codebase health, and understanding of failure modes.