Google published a study titled 'Agent Distraction': using data from hundreds of millions of coding agent steps, authors Sam Saccone and Madhuri Garikapati demonstrated that outdated inline comments like TODO(bug/...) and FIXME(bug/...) significantly derail agents from their tasks, resulting in measurable overhead in steps, tokens, and wait time in affected sessions. The company has already automated the cleanup of such comments in its codebases and recommends teams build 'agent-friendly codebases'.

image

What happened

On November 6, 2026, Google published the research paper 'Agent Distraction,' authored by Sam Saccone and Madhuri Garikapati. The analysis is built on data from hundreds of millions of coding agent steps. In 13% of file-reading sessions, agents encountered outdated inline comments TODO(bug/...) or FIXME(bug/...) with bug links. In 6% of cases, such contact triggered an action on a specific bug ID, with 83% of these actions being spontaneous—the developer had not asked to fix the bug. Of the spontaneous deviations, 41% ended in a dead end where the ticket was already closed, and 57% grew into scope creep. Multiple regression with fixed effects showed a statistically significant overhead: +70% steps, +77% tokens, +105% wait time with p<0.0001. Along with the results, Google described an already working intervention: automated agents scrape outdated links, thousands of dead TODOs were removed in six months, and human reviewers accepted such changes with an acceptance rate above 95%.

Context

The mechanism is inherent in how agents work: to perform a task, an agent reads files and takes the text around the code as working context, so an outdated bug link appears to it as an unclosed task and changes the action plan. Over time, some tickets are closed or links become outdated, but the comment remains in the code—classic technical debt, previously considered a cosmetic issue: its cost was measured by human annoyance, not machine money and time. The study translates this debt into agent economics for the first time: multiple regression with fixed effects controls for inter-project heterogeneity, making it difficult to explain the effect solely by repository differences, while the scale of data in the hundreds of millions of steps reflects real-world operation, not synthetic benchmarks. The background is that the share of code read and edited by agents is growing, and the repository has gained a new 'reader' for whom comment hygiene is not aesthetics but a direct cost of execution.

Why this matters for the industry

For the industry, this is the first quantitative measurement of the 'technical debt tax' when working with agents: outdated comments cease to be cosmetic and become a measurable expense in tokens, steps, and time. Consequently, cleanup task priorities change: projects that postponed removing dead TODOs now bear a specific cost in every agent run, and 'agent-friendly' codebase hygiene gains a ready-made ROI argument for engineering leads. Google explicitly recommends implementing presubmit analyzers and corpus scanners, and the company itself already operates production scrapers with an acceptance rate above 95%, meaning the approach is proven in practice, not just in theory. A product niche for hygiene tools opens up for vendors and startups: repository audits, auto-generation of cleanup PRs with mandatory human approval, simple presubmit hooks, and staleness metrics for TODO links in agent platforms. A likely direction in the coming months is presubmit checkers for dead bug links in CI and filtering of outdated comments in the agent context on the tool side, provided the effect is reproduced on third-party models and open repositories.

Why this matters for users

If you use coding agents like Gemini CLI, Copilot, or Cursor, dead TODO(bug/...)/FIXME(bug/...) in your repository may invisibly inflate task time and cost: the agent spends part of the session on a false goal instead of your task, sometimes drifting into scope creep or trying to 'fix' a long-closed bug. The practical step is simple and cheap—before mass agent runs, conduct an audit of the repository for outdated bug links: a grep script for TODO(bug/...)/FIXME(bug/...) patterns and a short review of the findings are sufficient. It makes sense to include an analyzer in presubmit checks that verifies the liveness of the ticket from the comment, so that new dead links do not accumulate in the codebase. The expected effect, according to the study data, is fewer unnecessary steps and token overconsumption in affected sessions and predictable agent behavior on a large corpus.

What is still unknown / limitations

The publication is hosted on the Google Developers corporate blog, and there is no open description of methods in the material: it is not specified which agents and models were observed, how 'spontaneous' runs were classified, and which covariates were included in the regression. All figures were obtained on Google's infrastructure, so direct transfer to other models, codebases, and programming languages cannot yet be considered proven. Until independent replications appear, specific percentages should be viewed as an estimate based on one company's data, not as a universal constant.

Sources

Author

Look at AI, editorial team