Trending storyDeveloping
Google study: stronger coding agents are more easily derailed by old TODO comments
1 report1 sourceupdated 6 hours ago
What happened
AI digest
On November 11, 2026, Computing Life reported on a Google study of more than 913,000 real coding sessions. It found that more capable models are more easily led astray by old TODO comments in source code. The study looked at coding agents that went off to chase historical comments on their own: in those sessions, a single conversation ran an average of 70% more steps, used 77% more tokens and waited 105% longer.
Written by AI from the coverage · updated 46 minutes ago
Coverage
Follow the reports to see the story from different sides.
Nov 12
- Computing Life · Share · Yage代码中的旧 TODO 注释为何更容易让强模型编码智能体跑偏
Google 对超过 91.3 万次真实编码会话的研究发现，能力更强的模型更容易被源码中的旧 TODO 注释带偏。智能体擅自追踪历史注释时，单次对话平均多执行 70% 的步骤、多消耗 77% 的 token，等待时间增加 105%。
Heat over time
Not enough continuous observations to draw a trend yet.