Agent Consistency, Thicker Interfaces, and Fraud at AI Startups

Three complementary signals today: Maggie Appleton makes a sharp case that agent interfaces are still too thin to support real collaborative planning, Thomas Wolf (via Hugging Face/IBM) surfaces the consistency problem in agent evals that's blocking production trust, and Stripe drops the most useful fraud data for AI startup PMs in months. Connecting thread: agents are capable enough to get the job done once, but we haven't built the infrastructure to make that repeatable and trustworthy at scale.