Continual learning for AI agents

Continual learning for AI agents occurs at three distinct layers: the model, the harness, and the context. Beyond just updating model weights, which faces challenges like catastrophic forgetting, learning can happen in the harness, which is the code and fixed instructions driving the agent. The context layer provides a third avenue for improvement by updating configurable instructions, skills, or memory, which can be done at the agent, user, or organizational level. Ultimately, all three learning flows are powered by analyzing execution traces of the agent's past behavior to inform improvements.