Google engineers analysed agent traces from thousands of developers and found about 13% of sessions exposed the agent to inline TODO or FIXME references. In 6% of those, the agent acted on the referenced bug, often without being asked, and 41% of such detours targeted tickets already closed. These sessions took 70% more steps and 77% more tokens. Google's automated cleaner removed thousands of obsolete references with over 95% reviewer acceptance, which suggests pruning stale comments is worth doing before rolling agents out.
