Agent Plasticity: Measuring Self-Improvement Through Experience

AI agents are increasingly being used in systems requiring them to self-diagnose and improve through experience, yet evaluations largely measure what they can do, not how effectively they learn. We study persistent self-improvement with frozen model weights, where agents learn by constructing and refining tools, skills, memories, and strategies inherited by future instances. Despite comparable learning opportunities, frontier models differ sharply in their ability to self-improve with experience, and gains on held-out in-distribution games often carry over only partly to harder, out-of-distribution games. We introduce agent plasticity, the efficiency of an agent in improving future performance as it gets more experience. Tracing failures through the improvement loop, we find that weaker agents often fail to reuse relevant artifacts, while highly plastic agents often fail despite reusing them, highlighting a shift in the bottleneck from reusing artifacts to building generalizable artifacts and applying them effectively. These results establish plasticity as a dimension of long-lived agent capability that static evaluations overlook.