Agent Plasticity Prices Agent Self-Improvement

This is a useful shift from asking how capable an agent is to asking how efficiently it becomes more capable.

• –Endpoint performance and learning efficiency can favor different models, making static leaderboards incomplete
• –Held-out evaluation exposes when agents overfit training interactions instead of learning transferable procedures
• –Low-plasticity agents often ignore relevant artifacts, while stronger agents can reuse artifacts that remain low-quality or poorly applied
• –The framework gives developers a practical way to compare reflection, memory, tool-building, and skill-generation loops
• –Cost estimates exclude deployment costs and depend on model pricing, saturation fits, and relatively small evaluation sets