Why do multi-agent system designers keep 'who does what' separate from 'what rules govern the group'?

This explores why designers of multi-agent systems separate what each agent does (its role) from the rules and shape of the group as a whole, and what that separation buys them. The short answer from the corpus is that roles describe individuals, but some of the most important constraints live *between* agents. The Gaia methodology made this point in 2003. Role models and interaction models can't express rules that span several agents, like "no reviewer may assess their own paper," or decisions about who is allowed to join the system at all. So organizational rules and structures have to be designed as their own first-class pieces, not inferred from the roles Should organizational rules be designed separately from agent roles?. Gaia also treats the environment as something to model explicitly, for the same reason: if you leave out anything that sits outside individual agents, the system you designed drifts from the system that actually runs Should multiagent systems model their environment explicitly?.

The MOISE+ model makes the split more concrete with three dimensions. The *structural* one covers roles and groups. The *functional* one covers the global plans the group is pursuing. The *deontic* one covers the permissions and obligations that connect who-you-are to what-you-must-do Can separating structure and function improve multiagent organization design?. The useful part is that roles become slots in a structure, and the rules are the glue between the slots and the work. MOISE+ even layers the links between roles: authority over a role requires a communication channel to it, and communication requires that the agents can represent each other. It checks this when the system is specified, not while it runs How do authority, communication, and acquaintance relate in organizations?. That is a design-level guarantee no single role description could give you.

The idea is now coming back in LLM-agent systems, sometimes under new names. IMACS splits team composition (Belbin-style roles), coordination mechanism (Mintzberg's organizational theory), and the algorithm that merges outputs into swappable layers, so you can change the team without rewriting the protocol Can multi-agent teams swap roles, coordination, and algorithms independently?. One finding is surprising: in a 25,000-task experiment, the best setup *fixed the structure* (an external ordering of who goes when) but *let agents choose their own roles*. It beat rigid hierarchies by 14% and fully self-organizing teams by 44%. Agents invented specialized roles and even sat out tasks they weren't good at Do self-organizing agent teams outperform rigid hierarchies?. So structure and role don't just differ in theory. You can set one firmly and leave the other loose, and that combination works best.

The flip side is that a rule written on paper does nothing unless the structure carries it out. In RACI experiments, labeling an agent "accountable" changed outcomes only when the workflow actually sent the deliverable through that agent Does accountability placement actually shape multi-agent outcomes?. Cognition's Walden Yan describes a related failure: parallel subagents each make unwritten assumptions their peers can't see, so they reach conflicting decisions. That's what happens when roles exist without shared structure or shared context to hold them together Why do parallel subagents produce conflicting design decisions?.

The lesson also applies to companies adopting AI. Deloitte argues that agentic AI fails when agents are slotted into old workflows. That amounts to assigning new roles without redesigning the organizational structure around them Does agentic AI succeed through process redesign or workflow layering?. The thing you might not have expected to want to know: a 20-year-old agent-engineering idea, that rules and structure are separate from roles, predicts much of what goes wrong with today's LLM agent teams and enterprise AI rollouts.