Agentic AI
How Stale Code Comments Are Tripping Up AI Coding Agents at Scale
Google's analysis reveals that outdated TODO comments in codebases are causing AI agents to waste time and resources chasing non-existent bugs.
Google Developers Blog (AI)
AI coding agents are getting distracted by obsolete bug references buried in codebases. According to research from Google Developers, abandoned TODO comments and FIXME markers—often left behind long after the underlying issue has been resolved—are misdirecting agents and inflating the computational cost of their work.
The Scale of the Problem#
When coding agents read through a codebase during planning and research, they treat these lingering comments as actionable tasks, not historical artifacts. Google's telemetry across millions of monthly agent sessions shows the scope:
- 13% of sessions expose agents to inline TODO(bug/...) or FIXME(bug/...) references
- 6% trigger secondary action—agents actually pursue the flagged issue
- 83% of those actions are unprompted, meaning the developer never asked the agent to investigate the bug
The real cost emerges when those detours lead nowhere. Cross-referencing unprompted investigations against live bug-tracking systems reveals two failure modes: 41% chase closed tickets that were already fixed or marked obsolete, while 57% trigger scope creep—agents expand the work scope to tackle open but unrelated issues without being asked.
The Hidden Tax on Agent Performance#
Using regression analysis across millions of sessions, Google quantified the overhead per engagement:
- 70% more steps to complete the task
- 77% more tokens consumed
- 105% more wait time for users
These numbers controlled for developer experience, codebase complexity, task type, model capability, and other variables—all statistically significant at p<0.0001.
Real-World Examples#
The research includes two case studies. In one, an agent investigating an unrelated bug stumbled on a TODO from 2023 and spent multiple steps researching it despite its fix date. In another, an agent encountered a referenced bug while doing legitimate work and unprompted began modifying code to "fix" the issue, only to be stopped when the human challenged the scope creep.

