AI NewsProductivityAnnouncement
Google research shows coding agents chase stale TODO comments
Google measured how often coding agents get distracted by stale TODO comments in a codebase. The paper is from Google's own engineers working on hundreds of millions of agent steps.

Image: Google
Why it mattersA stale TODO comment is free for a human to skip and expensive for an agent. Google measured 70 percent more steps and 77 percent more tokens per detour.
A stale TODO comment costs a human reader nothing to skip. It costs a coding agent extra steps, extra tokens, and extra wait time. Google published a paper on 6 November measuring how much. Sam Saccone is a senior staff software engineer. Madhuri Garikapati is a senior staff data scientist. They analysed hundreds of millions of agent steps across Google's codebases. They called the pattern "agent distraction".
What Google measured
Google looked at distributed traces from thousands of engineers running coding agents. The paper reports the numbers at each stage. In 13 percent of sessions the agent saw a TODO(bug/...) or FIXME(bug/...) comment in the files it was working on. In 6 percent of those cases the agent took a secondary action, and 83 percent of those actions were unprompted. Of the unprompted detours, 41 percent chased an already-closed ticket. 57 percent chased an open ticket that was outside the user's request.
Google then measured the cost of a detour. Google used a multi-way fixed-effects regression. Each unprompted bug detour added 70 percent more steps to the session. It added 77 percent more tokens. It added 105 percent more wait time. Those are per-session medians, not caps.
Why the comment is to blame
A TODO comment is a message a human wrote to another human. A coding agent reads it the same way a careful reviewer would, which means the agent treats it as context. 32 percent of the detours Google traced went to tickets already fixed and verified. 9 percent went to tickets marked "won't fix". Over half targeted tickets that were still open but outside the user's prompt. In every case the agent acted on information the code still carried. The code's owner had moved on.
What Google did about it
Google ran a scanner agent over its codebases for six months. The scanner reads a TODO and checks the referenced bug's status. If the bug is closed or dead, the scanner sends a pull request that removes the comment. Across thousands of such pull requests, human reviewers accepted the removal at a rate of over 95 percent. That is Google's number.
The paper recommends three practices. Run a presubmit analyser that catches a stale TODO(bug/...) before it merges. Run a scanner across the whole codebase to clean the back catalogue. Treat obsolete TODOs as a maintenance category, like dead code.
What this means for a team
Teams running Claude Code, Cursor, Codex or GitHub Copilot can read the paper as a budget line. A codebase with many old TODO comments pointing at a closed ticket tracker is paying extra tokens. Every detour costs 70 percent more per session. The fix is not an agent fix. It is cleaning up the signal the agent reads.
Source
Google Developers Blog, "Agent Distraction" by Sam Saccone and Madhuri Garikapati. Secondary coverage from InfoQ, "Google Finds That Technical Debt Can Derail Coding Agents through 'Agent Distraction'".
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.

