Nanjing University and University of Wollongong Propose Harness Continual Learning: Extending Continual Learning from Model Parameters to Agent Harness

Traditional continual learning focuses on updating model parameters along a sequence of tasks to avoid catastrophic forgetting. However, in the era of agentic AI, the behavior of deployed agents is determined not only by the foundation model but also by the harness surrounding it, including system prompts, interaction memory, reusable skills, external tools, workflows, and routing policies. Even without fine-tuning the model, updating these states can continuously change behavior.

Based on this, researchers from the Reasoning and Learning Research Group at Nanjing University and the University of Wollongong propose Harness Continual Learning (HCL), which for the first time extends the learning subject from model parameters to the harness, systematically defining and studying how the harness acquires new capabilities, preserves existing behaviors, and copes with forgetting along a sequence of tasks.

HCL is formally defined as follows: with the foundation model frozen, the agent updates the deployed harness along a sequence of interactive tasks, acquiring new behaviors while preserving past reliable behaviors as much as possible. Each update not only evaluates improvement on the current task but also asks whether it disrupts previously learned behaviors.