---
title: Agent Plasticity
type: vocabulary
url: "https://www.envisioning.com/vocab/agent-plasticity"
summary: A measure of how efficiently an AI agent converts experience into lasting held-out performance gains, via inherited artifacts rather than weight updates.
year: 2026
generality: 0.40
---

# Agent Plasticity

A measure of how efficiently an AI agent converts experience into lasting held-out performance gains, via inherited artifacts rather than weight updates.
Agent plasticity is a measure of how efficiently an AI agent converts experience into lasting improvement on held-out tasks, without updating its model weights. Harman Singh, Anton Bakhtin, Rulin Shao, Gabriel Synnaeve, Ilia Kulikov, Rob Fergus, Sanjeev Arora, Kurt Keutzer, Jason Weston, Anuj Mahajan, and Anirudh Goyal, at Meta Superintelligence Labs, UC Berkeley, the University of Washington, and Princeton, introduced the term in the October 2026 paper "Agent Plasticity: Measuring Self-Improvement Through Experience" (arXiv:2610.08902). It targets a gap in standard agent benchmarks: they score what a model can do at one point in time rather than how effectively it learns from its own experience.

In the paper's setup, a model's weights stay frozen. Each new instance of the agent inherits an "inventory" of artifacts, such as written tools, skills, and memory notes, left behind by earlier instances. After every round of training games, a fresh instance reflects on the outcome and proposes edits to the inventory. These edits must pass structural and executable checks before the next instance inherits them. Performance is then scored on held-out games that the reflection step never saw, split into in-distribution and out-of-distribution opponents. Agent plasticity is defined as the held-out score gained divided by the learning cost (the API spend on training and reflection) up to a chosen checkpoint. A headline variant, Psat, stops the measurement at the point where a fitted learning curve reaches 90 percent of its saturation.

The authors tested frontier models on chess, Go, Hex, and NetHack and found that endpoint capability and learning efficiency diverge. Claude Fable 5 reached the highest final held-out score but took about five times the learning cost to saturate compared with GPT-5.6 Sol, which had the highest plasticity score despite a lower ceiling. Some models, including Claude Opus 4.8 and GPT-5.6 Luna, showed little or no persistent improvement across rounds. The paper also traced failures to artifact use: models with low plasticity often skipped relevant artifacts entirely, while some high-reuse models still failed on a majority of their remaining mistakes even while using a relevant artifact. That pattern points to limits in artifact quality or application rather than retrieval.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/agent-plasticity)