The big-picture lesson

The most important takeaway is the one stated up front: improving a multi-agent system is mostly about better orchestration, not bigger models or more tokens. Many failures come from agents operating on incorrect assumptions, ignoring peer input, or failing to verify their outputs, all of which are design problems. The article's broader contribution is to move the whole conversation from anecdote ("multi-agent systems are flaky") to diagnosis ("here is exactly which of 14 failure modes is hurting your system, and here is how to measure improvement").