Arkiti pushed me to think beyond a single model call. A multi-agent architecture is useful when different parts of the product need different responsibilities, context, and decision boundaries.
Agents need jobs, not personalities
The useful question is not "how many agents can we add?" It is "which responsibility deserves its own reasoning loop?" Agents should reduce complexity, not create a maze.
Coordination is the product
Once multiple agents exist, orchestration becomes the real system. You need clear handoffs, shared context, failure states, and a way to inspect what happened.
The practical lesson
Multi-agent systems work best when they are boring at the edges: typed inputs, predictable outputs, defensive logging, and constrained responsibilities. The intelligence can be flexible, but the interfaces around it should be stable.