It appears like the 'self-improving' here just means modifying the agent's prompt/context? And not actually changing any of the weights/architecture of a model. I feel like this kind of self-improvement has some hard limits on how much it can improve.