Fascinating set of insights into how far an AI agent can go, and how it breaks down. The other thing I find refreshingly honest, is the tone of the company who make the thing, openly saying "we have no idea why it did this". That bedrock fact underlies all LLM development, and any "AI" company who claims to truly, properly, understand their product is lying.