Tip: Agents Don't Refactor — Make Cleanup an Explicit Gate, Not a Hope

paddo.dev·medium signal

paddo.dev's April 23 post argues that autonomous agents primarily add new code rather than cleaning up existing code, and that expecting agents to refactor is a category error. The actionable technique: implement cleanup as an explicit, separate step in your agent workflow with its own verification criteria — not as a hoped-for side effect of feature work. This matches the observable pattern in Claude Code and Cursor where agents solve the stated problem but accumulate technical debt if not explicitly directed otherwise.