LLM Inference Interview Questions, issue 25, Aug 25, 2026
The Irreversible Action Trap
Senior AI Engineer interview at Stripe, and the interviewer asks:
“Your LATS agent crushes HotpotQA. Now point it at a checkout flow where one of the actions is ‘submit payment.’ How do you keep tree search from becoming a production incident?”
Don’t say: “I’d add a human-in-the-loop approval step.”
Why letting MCTS agents explore real-world tasks quietly triggers production disasters, and how partitioning your action space saves the system, but shifts your engineering bottleneck to sim-to-real
The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.