Your Agent Aced the Task. Will It Do It Again?

AI agents often exhibit unreliable performance, succeeding on a task once but failing on subsequent identical attempts. Standard benchmarks that report average success rates hide this variability, creating a significant "consistency gap." A new diagnostic tool, the Consistency Analyzer, identifies flip-prone decision points in an agent's reasoning process by resampling its past actions. These diagnoses are then used by the ALTK-Evolve system to create consistency guidelines that are injected at inference time, which demonstrably improves an agent's reliability and reduces the performance gap without harming overall accuracy.