Unstated Rules: Why AI Agents Over-Apply Changes

The true bottleneck in AI-assisted development is not agent capability but the gap between what a codebase implicitly knows and what's explicitly accessible to agents. This analysis of 767 sessions reveals that 60% correction rates stem from agents over-applying changes without understanding unstated rules about naming conventions, scope boundaries, and architectural patterns. As codebases grow more complex, developers face an increasing cost of teaching agents these implicit constraints—making the correction loop a systematic tax on knowledge transfer.