AI code refactoring uses large language models to improve code structure without changing external behavior, targeting issues like poor naming, duplication, and complexity. Two main approaches exist: inline refactoring (small, editor-level suggestions alongside a developer) and agentic refactoring (autonomous, goal-driven changes across an entire codebase). Agentic refactoring works as a loop: planning, reading files, searching for related code, reporting findings, getting human approval, patching, and verifying via tests and builds. Guard rails include human approval checkpoints, rollback snapshots, test/build verification, and optionally deterministic methods like abstract syntax trees instead of probabilistic LLM guesses. The loop can integrate into CI/CD pipelines, and reinforcement learning from accepted/rejected fixes improves suggestions over time.