AI Planning Agent Flagged 96 of 97 High-Stakes Goals — and That Was Intentional

A developer building an AI planning agent called PlannerCritic found that 96 out of 97 strict goals were escalated rather than approved during field testing. Initially appearing to be a failure, the high refusal rate was later reframed as a sign of correct behavior in a high-stakes planning system. The developer argues that a plausible but flawed plan is more dangerous than a clean refusal, since hidden assumptions in generated plans can mislead human reviewers before errors are caught. PlannerCritic was designed to produce only three terminal states — approved, escalated, or blocked — rather than maximizing plan output. The experience prompted a reassessment of how AI planning agents should be evaluated, shifting focus from approval rates to the quality and honesty of refusals.