At work in an AI heavy environment I sometimes let the agent just try to implement some spec not with the goal of actually using that but to use it as an input for critique, especially when it gets stuck and tries increasingly zany hotfixes. Learning from failure is a thing but learning from AI failure is a thing, too