> The agent cannot learn from its mistakes. The agent will never produce any output which will help you invoke future agents more safely

That is not entirely true:

Given that more and more LLM providers are sneaking in "we'll train on your prompts now" opt-outs, you deleting your database (and the agent producing repenting output) can reduce the chance that it'll delete my database in the future.