What happened
When the same AI model solves the same problem multiple times, it makes the same mistakes repeatedly instead of varying its approach. This means you can't improve accuracy just by running the model again and picking the best answer—the model needs to actually learn a better strategy.
Why it matters
Before deploying AI agents in production systems that make real decisions, companies need to know whether failures are fixable through redundancy (run it again, pick the best output) or structural (the model itself is wrong). This paper shows redundancy doesn't help when an agent confidently misunderstands the problem the same way every time.