Stale instruction files add hidden cost to coding agents
The point
Researchers studied three real stale instructions in active open source repos: a test command, a code location, and a component procedure. Agents fixed 22 tasks under stale, corrected, and removed versions, across 732 trajectories from two models. They followed stale guidance when relevant, even when correct information was available elsewhere.
Why it matters
For builders, current AGENTS.md and CLAUDE.md files keep agent runs cheaper and more reliable.
Key details
- Claude Sonnet 5.5 needed about two extra tool calls per run on the stale test command
- No agent repaired the stale instruction file
Original ~4 min → 25 sec here Source: arXiv cs.AI, Nov 9, 2026