Stalls, disagreement, and duplicated work
Once agents depend on each other, new failure modes appear that don't exist in a single-agent system. Agents can get stuck waiting on each other, for example a supervisor waiting on a worker that itself is waiting on a result the supervisor was supposed to provide, a deadlock. Agents can also disagree, a critique agent might reject a writing agent's output, but if there's no defined process for what happens next, the system can loop indefinitely with no path to a final answer.
A subtler problem is duplicated work: when two agents' responsibilities overlap even slightly, both might end up doing the same research or producing overlapping drafts, wasting time and tokens and sometimes producing conflicting outputs that a downstream agent has to reconcile. All three problems trace back to the same root cause, unclear ownership of who is responsible for what and when.