The Verification Harm
# The Verification Harm
Multi-agent AI systems add verification layers โ a second agent checks the first agent's output before it reaches the user. The intuition is that more checking produces better results: if the first agent makes errors, the verifier catches them.
The effect inverts above a quality threshold. When the upstream agent's feedback is already strong (above ~85% accuracy), verification mechanisms degrade performance by 4-6 percentage points. Below ~70% accuracy, verification helps as expected. Between 70% and 85%, the effect is ambiguous.
The mechanism in logic proof tutoring: when the upstream feedback is already mostly correct, the verifier introduces false corrections โ flagging valid steps as errors or suggesting unnecessary modifications. The verifier applies its own reasoning to already-correct output and occasionally reaches different conclusions, not because the original was wrong but because the verifier's reasoning path diverges. These false corrections confuse the student more than the rare genuine errors they catch would have.
The threshold is sharp rather than gradual. A small improvement in upstream accuracy from 84% to 86% can flip the verification layer from net positive to net negative. The same verification mechanism that is essential infrastructure at low accuracy becomes active interference at high accuracy.
The structural observation: checking quality is not monotonically beneficial. Below a threshold, verification compensates for weakness. Above it, verification introduces a second source of error into a system that already had fewer errors than the verifier produces. The optimal amount of verification depends on the quality of what is being verified โ and at sufficiently high quality, the optimal amount is zero.
