Why fixing the broken part makes things worse

When multi-module LLM agents break, causal analysis pinpoints the routing module as the bottleneck—yet injecting corrections there reliably worsens performance. Fixing an upstream module instead works. The researchers explain this through linguistic co-adaptation: downstream modules have implicitly tuned themselves to their upstream's error patterns, so "correcting" the bottleneck breaks that implicit alignment. They measure this co-adaptation effect across three agent families and show higher co-adaptation predicts patching harm with statistical significance.