Agent Plasticity Measures How Efficiently AI Agents Improve Through Experience

Summary

The paper argues that standard agent evaluations capture capability at a fixed point but overlook how well an agent learns from experience. It introduces “agent plasticity,” defined as the efficiency with which an agent converts past experience into improved future performance on held-out interactions. In a controlled setting, agents amortize experience into reusable artifacts that are inherited by later instances, and performance is measured at successive checkpoints on both training and held-out interactions while learning costs are accounted for. Across multiple environments, frontier models show sharply different improvement trajectories despite comparable opportunities to learn: some achieve substantial, persistent gains, while others stay near or below their starting performance. Improvements on the training regime transfer only partially to out-of-distribution conditions. The study also finds that final capability and improvement efficiency can diverge, so the best endpoint performer is not necessarily the fastest or most efficient learner. Failure analysis identifies different bottlenecks: low-plasticity agents may fail to reuse relevant artifacts, while agents that reuse them can still fail because of artifact quality, generalization, or application problems. The authors conclude that evaluating self-improving agents requires measuring how they become better, not just what they can do at one moment.