An agent can improve without retraining the model.

Once agents rewrite their own prompts, memory, and routing, every update becomes a behavioral change.

As agents keep rewriting prompts, memories, skills, and routing rules, those updates become part of what the system has learned.

Fixing today's failure can quietly break something that worked yesterday.

The paper calls this harness-level forgetting: the model stays frozen, but the behavior around it keeps changing.

Their Harness Continual Learning (HCL) framework puts every proposed harness update behind a gate.

For evolving agents, prompt, memory, skill, and routing changes should be handled like code changes: regression-test them before they become persistent.

– arxiv. org/abs/2608.19013

Title: "Harness Continual Learning: Continual Adaptation Beyond Model Parameters"