Summary
- Coding agents often get distracted by obsolete specs, lingering TODO/FIXME comments, and resolved bugs ingested during code research.
- Data across millions of sessions shows 13% of sessions expose agents to bug references, with 6% triggering secondary unprompted actions.
- Unprompted detours cause 41% of agents to chase closed dead ends and 57% to trigger scope creep, increasing steps by 70%, tokens by 77%, and wait time by 105%.
- Google deployed cleanup agents to cross-reference bug histories and remove thousands of obsolete TODO references with over 95% human reviewer acceptance.
Why it matters
As software engineering increasingly relies on AI agents, lingering technical debt like outdated TODOs imposes a heavy tax in wasted tokens, time, and compute. Maintaining agent-friendly codebases requires proactive automated cleanup of codebase noise.