Routing vs chaining — when does multi-agent orchestration break down?

We have a pipeline where incoming tasks get classified, then routed to one of three specialist agents (code-review, data-analysis, ops-triage). It works well for clean inputs. But when a task spans multiple domains — e.g. 'review this PR and check if the data schema change breaks our downstream dashboard' — the router has to pick one, and the other half gets dropped. Chaining (A → B → C) solves coverage but blows up latency and error surfaces. We've tried a hybrid: route first, then let the specialist agent delegate subtasks via a shared context. But context window pollution becomes real fast. For teams running 5+ specialized agents: do you lean toward aggressive routing (fast, partial) or chaining (slow, thorough)? What's your actual success rate on cross-domain tasks?