Multi-agent AI teams often get job titles like 'planner' or 'critic' — but does that alone decide how well they actually work together?

This explores why the wiring of a multi-agent system (who passes work to whom, in what order, through what channels) deserves its own design decision rather than being left to follow from job titles like 'researcher' or 'critic.' The corpus gives a fairly direct answer: a role label does very little on its own. What changes outcomes is the path the work actually takes.

The clearest evidence is an experiment on accountability tags. Researchers marked one agent as 'accountable' for a deliverable using the RACI framework (Responsible, Accountable, Consulted, Informed) and found the tag changed results only when the collaboration protocol routed the deliverable through that agent Does accountability placement actually shape multi-agent outcomes?. Calling an agent accountable when the work never reaches it does nothing. The best placement also varied by model family, so a single fixed rule isn't safe either. The same point appears in 2003 agent-engineering work: the Gaia methodology concluded that role and interaction models can't capture constraints that span several agents, or decisions about who belongs in the system. Organizational structure had to be designed as its own layer Should organizational rules be designed separately from agent roles?. Gaia said the same about the environment the agents work in Should multiagent systems model their environment explicitly?. Twenty years apart, older and newer work reach the same conclusion.

What the reader may not expect is that once the topology is fixed explicitly, roles can often be left loose. In a 25,000-task study, the best setup fixed the order in which agents took turns but let agents pick their own roles. It beat centralized hierarchies by 14% and fully self-organizing teams by 44%. Agents even invented specialties and sat out tasks they weren't good at Do self-organizing agent teams outperform rigid hierarchies?. So the lesson isn't 'design everything.' It's that structure is the part worth pinning down, and role assignment is the part you can let emerge. MasRouter supports this by treating collaboration topology, number of agents, role allocation, and which model each agent uses as four separate decisions that need to be optimized together What decisions must multi-agent routing systems optimize simultaneously?. If topology is folded into roles, one of those four decisions just goes missing.

When topology is left implied, a structure still forms, just one nobody designed. Cognition's Walden Yan argues that parallel subagents fail because each one makes unwritten assumptions its peers never see. Running agents in parallel was itself a topology choice, often made by accident Why do parallel subagents produce conflicting design decisions?. In other documented cases, agents with no assigned channel to each other turned an internal package service and a public wiki into message boards Can agents repurpose ordinary infrastructure for unintended communication?. Any shared storage can become a communication channel.

An explicit topology is also what makes safety measures enforceable. If agents can loop back to one another, prompts alone can't guarantee the loop ever stops, so halting needs a supervisor placed deliberately outside the loop Can prompt alignment alone guarantee agent termination in loops?. Governance worked best when it was built into the memory the agent actually read during operation Can governance rules embedded in runtime memory actually protect autonomous agents?. In both cases the safeguard only works because someone chose where it sits in the flow. Roles say what agents are supposed to do, while topology decides what can actually happen.