- Google analysed hundreds of millions of coding-agent steps taken by its own engineers.
- When an agent chased an old bug note on its own, the session took 70% more steps, 77% more tokens and 105% more waiting.
In short: Google engineers looked at what happens when a coding agent runs into an old to-do note in the code that points to a bug ticket. A post on the Google Developers Blog from 6 November 2026 shows that the agent usually skips such notes, but when it does act on one, nobody asked it to most of the time, and the whole job gets much longer.
What Google measured
The analysis covers agent traces from thousands of Google engineers and hundreds of millions of individual steps:
- 13% of sessions in which the agent read files ran into a note like this.
- In 94% of encounters the agent ignored it. In 6% it did something about the bug, and in 83% of those cases nobody had asked it to.
- 41% of these unprompted detours chased tickets that were already closed: fixed (32%) or abandoned as obsolete (9%).
- 57% involved bugs that were still open. That is classic scope creep: in 39% of those cases the agent edited code nobody asked it to touch.
After controlling for the developer, code complexity, model tier and more, such a detour meant 70% more steps, 77% more tokens and 105% more waiting per session.
What Google did about it
The company built agents that combed through its code, found notes pointing to closed tickets and checked the history of the bug and the files, sometimes going back years. They proposed deleting a note only once they confirmed it was obsolete. Google says this removed thousands of notes over the last six months, with human reviewers accepting more than 95% of the changes.
How to read this
These are numbers from inside Google, from its own tools and codebase, so the ratios in your project may differ. The mechanism is simple and applies to any codebase, though: stale notes, old specs and outdated docs are instructions to an agent, not background noise. Before you pay for a stronger model, check what the agent reads along the way. I write more about working with agents in the Coding agents hub.