Google: Technical Debt Derails AI Coding Agents via 'Agent Distraction'
Google researchers found that AI coding agents are frequently derailed by outdated technical debt like stale TODO comments and obsolete bug references, a phenomenon called "agent distraction" that wastes resources and can cause incorrect logic. In a study of hundreds of millions of agent steps, Google discovered that roughly 13% of sessions exposed agents to such references, leading to unrequested detours in 6% of cases. These side-quests added 70% more steps, 77% more tokens, and 105% more wait time per session, with many targeting already-closed bugs or out-of-scope issues. To counter this, Google developed scanner agents that clean up obsolete TODOs, removing thousands of references over six months with over 95% human reviewer acceptance. The researchers recommend presubmit analyzers and automated cleanup tools to mitigate distraction and improve AI coding agent effectiveness.