Giving an AI agent a fancy job title like 'manager' does nothing unless the system actually routes work through it.

This explores whether labeling agents with organizational roles changes anything on its own, or whether the routing protocol, meaning the control flow that decides which agent sees which work and when, is what gives a role its force. The clearest answer in the corpus is a little humbling: roles do very little unless the routing makes them real. Ablation studies on RACI-style tags (Responsible, Accountable, Consulted, Informed) found that marking an agent as "accountable" changed outcomes only when the collaboration protocol actually passed the deliverable through that agent Does accountability placement actually shape multi-agent outcomes?. A role the protocol never routes through is just a label. Even when it does work, the best place to put accountability varied by model family, so you can't copy one org chart across every model and expect the same results.

This matches an idea organizational theorists in multi-agent systems reached about twenty years before LLMs. The Gaia methodology (2003) argued that role definitions alone can't capture rules that span several agents, so organizational rules and structures need to be designed as separate things in their own right Should organizational rules be designed separately from agent roles?. MOISE+ went further and split an organization into three layers: structure (roles and groups), function (shared plans), and deontic rules (who is permitted or obliged to do what), with the deontic layer connecting the other two Can separating structure and function improve multiagent organization design?. Read alongside the RACI result, the older work suggests that today's "routing protocol" is doing the job those deontic rules did: it is what turns a role into an obligation.

If roles and routing are separate concerns, can you build them separately? IMACS says yes. It treats team structure (Belbin-style roles), coordination mechanism (drawn from Mintzberg's organizational theory), and the algorithm that merges outputs as swappable layers, so you can change the team without rewriting the protocol Can multi-agent teams swap roles, coordination, and algorithms independently?. MasRouter comes at it from the other direction and treats role allocation as one of four routing decisions to optimize together, alongside collaboration topology, number of agents, and which LLM each agent runs on What decisions must multi-agent routing systems optimize simultaneously?. One view keeps roles and routing apart for modularity. The other merges them for performance. Both agree the two are tightly linked.

The most surprising finding is that agents may not need assigned roles at all if the protocol is shaped well. In a 25,000-task experiment, a hybrid setup worked best: a fixed external order of turns, with agents choosing their own roles inside it. It beat centralized hierarchies by 14% and fully self-organizing teams by 44%. Agents invented specialized roles on their own and sat out tasks they weren't good at Do self-organizing agent teams outperform rigid hierarchies?. So the protocol can act as the scaffold that roles grow out of, rather than something that carries out an org chart you designed in advance. In open ecosystems, versioned capability vectors push this further by letting agents be discovered by what they can do instead of by manually assigned roles Can semantic capability vectors replace manual agent routing?.

The gap is what happens when routing crosses organizational boundaries. Once one company's agent hands work to another company's agent, nobody is clearly named as owner of the rules that should govern that path, and operator, organization, regulator, and standards body may each want different things Who enforces invariants when agents cross organizational boundaries?. Taken with the RACI result, this points to a real risk: if accountability only counts where routing passes through it, a handoff between organizations can silently route around the agent that was supposed to be accountable.