Google: stale TODO comments derail 1 in 7 coding agent sessions
rseroter · x · 2026-11-07
Google's 'Agent Distraction' study, based on distributed traces across thousands of engineers and hundreds of millions of agent steps: 13% of file-reading sessions expose coding agents to inline TODO(bug/...) comments, and 6% trigger action. Of agents acting on a TODO, 83% do so unprompted — 60% only look up the bug, 40% edit code. The referenced bugs are 57% still open, 32% already fixed, 9% obsolete. Best case it wastes tokens and time; worst case agents ground on stale falsehoods. Recommendation: clean up stale TODOs and outdated docs, and use the TODO-vs-ticket-status cross-reference to measure distraction at scale.
More from coding & agent
- Pair programming with Codex listening in: design debates yield simpler agent code — angadsg · 2026-11-07
- X launches new Livestreaming API routes, old ones EOL Jan 1, 2027 — Kyrannio · 2026-11-07
- Agents are like Memento: your system prompt is their tattoo, most context never reaches them — brandon_galang · 2026-11-07
- Hacker gets Hermes Agent running on a Nintendo 3DS with voice input via custom homebrew — alexcovo_eth · 2026-11-07
- Developer's bar for AI code: if you can't explain it in review, the PR isn't done — therealdanvega · 2026-11-07
- Claude Code cloud sessions now auto-detect dependencies and securely store API keys — ClaudeDevs · 2026-11-07