Agent Plasticity: Measuring How AI Agents Learn from Experience

Introduction

A high score on a static benchmark does not necessarily mean that an AI agent can learn well. In real deployments, agents encounter failures, changing tasks, and unfamiliar situations. They must diagnose what went wrong, preserve useful lessons, and make those lessons available to future runs. The paper Agent Plasticity: Measuring Self-Improvement Through Experience focuses on this missing dimension of evaluation: how reliably and efficiently an agent becomes better over time.

From endpoint performance to learning dynamics

Most agent benchmarks report capability at a fixed point. Such scores are useful, but they do not answer whether an agent can improve after experience, how much interaction is required, or whether improvements survive outside the environment in which they were learned. The study creates a controlled setting in which agents convert previous interactions into reusable artifacts, including tools, skills, memory, and textual knowledge. Later instances inherit these artifacts and use them in subsequent tasks.

The authors introduce “agent plasticity,” describing the efficiency with which an agent converts learning cost into gains on future held-out performance. Evaluation takes place at multiple checkpoints and compares results on both the training distribution and unseen interactions. This makes it possible to distinguish genuine transfer from narrow adaptation to familiar examples.

Key findings

• Frontier models follow sharply different improvement trajectories despite receiving comparable opportunities to learn. Some produce substantial and persistent gains, while others remain close to their starting point or even fall below it.
• The strongest endpoint agent is not necessarily the most efficient learner. A system can require more experience to reach a high final level, while another gains more from each unit of learning cost but ends at a lower ceiling.
• Improvements within the training regime often transfer only partially to out-of-distribution conditions. An artifact may capture a local tactic rather than a broadly useful solution principle.
• Self-improvement is more than retrieving old experience. Agents must create artifacts that contain actionable knowledge and then apply them appropriately. Failures in retrieval, artifact production, or execution can break the improvement loop.

Why it matters

The study points toward a broader evaluation stack for long-running agents. In addition to endpoint scores, developers should track learning curves, cost-normalized gains, held-out generalization, and the durability of improvements. Simply allowing an agent to run for longer does not guarantee that it will become more capable; experience must be selected, compressed, and reused effectively.

The framework is also useful for diagnosing systems. If performance does not improve, the cause may be poor experience retrieval, low-quality artifacts, incorrect timing, or unreliable execution rather than a complete inability to learn. Separating these stages can guide both architecture design and training procedures.

As agents become persistent software systems rather than one-shot model calls, their ability to improve through experience will become a core product property. Agent plasticity offers a way to ask a more practical question than “How capable is the agent now?”: “How much better can it become, at what cost, and does that improvement hold in situations it has not seen?”

Source: Hugging Face Daily Papers