Researchers propose a new approach to continual learning called Harness Continual Learning (HCL). Unlike traditional model-centric approaches, HCL focuses on adapting the 'harness' of an agent, which includes prompts, memories, tools, skills, and routing rules. This harness evolves around a frozen foundation model, allowing for continuous improvement without disrupting previous behavior. The authors introduce four components to implement HCL: Task Interface, Experience Memory, Capability Map, and Adaptive Router. They also propose a Continual Optimizer and Evaluator to manage harness evolution. Experiments demonstrate the effectiveness of HCL in various tasks, including textual reasoning and multimodal perception. --- Why it matters: This work matters because it offers a new perspective on continual learning, which is essential for developing AI agents that can adapt to changing environments without forgetting previous knowledge. By evolving the harness around a frozen model, HCL enables continuous improvement while retaining previously acquired behavior. Source: https://arxiv.org/abs/2608.19013