Accurate Failure Prediction in Agents Does Not Imply Effective Failure Prevention

Rakshith Vasudev, Melisa Russak, Dan Bikel +1

Proactive interventions by LLM critic models are often assumed to improve reliability, yet their effects at deployment time are poorly understood. We show that a binary LLM critic…