Giving AI agents different job titles sounds smart, but does it actually stop them from doing the same work twice?

This explores whether giving agents different roles in their prompts is enough, on its own, to stop them from duplicating work and repeating information. The corpus has no study that measures redundancy directly. What it does have points one way: role labels help, but they rarely do the work alone. What changes outcomes is the structure around the roles: who passes what to whom, and in what format.

The clearest evidence comes from studies that test role assignments directly. In one, accountability tags (who is "responsible" or "accountable" for a deliverable) changed outcomes only when the collaboration protocol actually routed the work through the agent holding that tag Does accountability placement actually shape multi-agent outcomes?. A role written into a prompt is close to decoration until the system's control flow gives it consequences. The best placement also varied by model family, so a role design that works for one model may not carry over to another. IMACS reaches a related conclusion from the design side. It treats team roles, the coordination mechanism, and the method for merging agents' outputs as three separate, swappable layers Can multi-agent teams swap roles, coordination, and algorithms independently?. On that view, roles are one layer among three. They cannot also do the job of the coordination layer.

For redundancy itself, the strongest lead is MetaGPT. Its agents coordinate better when they produce standardized documents and pull what they need from a shared workspace, instead of chatting back and forth Does structured artifact sharing outperform conversational coordination?. Conversation is where redundancy builds up: agents restate, summarize and re-explain to each other. A fixed artifact format limits what each role is allowed to contribute. That suggests redundancy is mainly a problem of communication format, not of role assignment. Large-scale coordination tests point the same way. Agents tend to accept whatever their neighbors tell them without checking it, so information and errors spread through the network, and coordination gets worse as the network grows Why do multi-agent systems fail to coordinate at scale?. A role prompt doesn't change that habit.

One point is easy to miss. If roles are "just prompts," you may not need multiple agents at all. A single model switching between personas can reproduce much of what multi-agent debate achieves Can branching prompts replicate what multi-agent systems do?. That cuts both ways. Role prompting is powerful, but agents that differ only in role may share the same underlying knowledge, so their "different perspectives" overlap more than they appear to. Work on social simulation shows models look competent when one model controls every participant, and fail once agents hold genuinely private information Why do LLMs fail when simulating agents with private information?. Real division of labor may depend on agents actually knowing or seeing different things, not on being told they are different. The case for multi-agent systems rests on that kind of separation: different expertise, parallel work, and independent checking Do single agents always hit organizational limits?.

So the corpus's answer is "probably not alone." Role prompts set up the division of labor. Routing, shared structured artifacts and real differences in what each agent knows are what stop agents from stepping on each other.