Google measured millions of internal agent sessions and found that stale TODO comments in code lead agents into detours; sessions with such unrequested detours take 70% more steps and 77% more tokens than sessions without them.

Part of agents' cost comes from codebase hygiene rather than the model, though this is Google's measurement of Google's internal codebase and internal agents, so it extrapolates only so far. Google's developer blog wrote on November 6 that 13% of sessions that read files hit a TODO comment linked to a bug-tracker ticket. When agents hit such a TODO, they acted on the ticket 6% of the time. Developers hadn't asked for 83% of those actions. Of the unrequested detours, 41% chased tickets that were already closed. After a regression controlling for developer, codebase complexity, task type and model tier, sessions with such detours took 70% more steps, 77% more tokens and 105% more wait time than sessions without. That is a correlation after regression controls, not causation. Google has deployed an agent that cleans up stale TODOs; over the past six months it deleted thousands of them, and human reviewers accepted more than 95% of the changes. (Google for Developers, 2026-11-06). How we checked →
Source and checks
This item ran in the brief of Nov 7, 2026, alongside 13 other items. How we checked it, the quote checks and what was still missing that day are on that issue's How we checked page.
Cite this item
SecondSource (2026). "Google measured millions of internal agent sessions and found that stale TODO comments in code lead agents into detours; sessions with such unrequested detours take 70% more steps and 77% more tokens than sessions without them.." https://secondsource.io/en/research/2026-11-07-2/. Accessed access date.