We Let Claude Code Refactor Our 200K-Line Java Monolith. Here's the Honest Result.

An engineering team recently tested the capabilities of the AI agent Claude Code by tasking it with refactoring a 200,000-line legacy Java monolith. Over six weeks, the team shifted from fully autonomous overnight runs to a human-gated hybrid model. While the AI excelled at mechanical tasks like moving classes and modernizing Java syntax, it struggled with architectural nuances, such as maintaining transactional boundaries and preventing test coverage erosion. The experiment concluded that while AI can significantly accelerate refactoring—completing work in six weeks that would have taken months manually—it requires strict human oversight, clear constraints via a detailed CLAUDE.md file, and automated verification scripts. The team successfully migrated the codebase with zero production bugs, demonstrating that AI is most effective when humans define narrow, specific boundaries for autonomous tasks rather than delegating broad architectural decisions.