Google Finds AI Agents Chase Stale Code Tasks
Google's research found coding agents often pursue unrequested fixes triggered by outdated TODO references left in codebases.
Updated on Nov. 10, 2026 in Artificial Intelligence
Google published findings that AI coding agents frequently detour into unrequested work when they encounter leftover TODO or FIXME references in code. The detours were found to add substantial steps, tokens and wait time to each session.
Why it matters
Agents that pick up obsolete references can waste compute, slow developer workflows and act on closed or out-of-scope tickets. The findings point to a practical cleanup problem as AI coding tools spread through software development worldwide.
The analysis drew on distributed traces from thousands of engineers and hundreds of millions of agent steps, with 41% of unprompted detours targeting already closed tickets and 57% targeting open bugs outside the requested scope. Google's todo cleaner removed thousands of obsolete references over six months with over 95% human reviewer acceptance, but the study does not say how broadly these results generalize beyond Google's internal codebase.
The players
Google is the technology company whose research team published the study on AI coding agent distraction and built the todo cleaner scanner agents.
The details
Google found that agents frequently treated leftover TODO(bug/...) comments as instructions, even when the prompt did not mention them. To address the problem, Google built scanner agents that inspected git history to confirm a TODO and its intent were obsolete before proposing removal.
Timeline
November 10, 2026: Google publishes its agent distraction findings.
Over the preceding six months: the todo cleaner removes thousands of obsolete references.
The Tech Race
The findings extend the long-standing software convention of tagging bugs with TODO(bug/...) comments, which were written for human readers but are now ingested as context by autonomous coding agents. As AI coding tools compete to handle more of the development workflow, keeping repositories free of stale references is becoming a measurable factor in agent performance.
Developers using AI coding agents may see slower sessions and higher token costs when old TODO comments in their repositories pull the agent off task. Cleaning out obsolete references before running agents could reduce wasted steps and wait time.
The takeaway
Keeping code comments current is becoming an operational concern, since stale references can now steer autonomous tools toward work nobody asked for. Teams adopting AI coding agents may benefit from periodic audits of TODO and FIXME markers.
Further reading
Explore more reporting in the Artificial Intelligence section for context on how AI tools are changing software development.
