Before building a team of AI agents, there is a more fundamental architecture question: why should the second agent exist at all? Another capability does not automatically require another autonomous boundary.
The same work may belong in a tool, function, instruction, workflow node, or internal role.
The stronger test is whether separation preserves something a simpler architecture cannot: independent evidence, fault isolation, authority, state, accountability, or recovery. If that boundary stops creating more value than coordination and governance cost, the agent should not become permanent infrastructure.
Multi-agent architecture should be qualified continuously (not assumed at design time).