Project Deal: Our Claude-Run Marketplace Experiment – Anthropic

fascinating experiment. i've been building some local multi-agent setups recently, and the biggest headache is always when agents silently fail on edge cases or get stuck in weird loops.

trusting an agent with actual marketplace transactions still feels a bit terrifying tbh. did they detail how they built the fallback guardrails if claude confidently decides to do something completely irrational with the pricing?