The easiest way to make a multi-agent system look serious is to give the agents roles. One is the sceptical regulator. One is the product owner. One is the security reviewer. One is the customer advocate. The transcript instantly feels richer. Different voices appear to disagree. The system looks less like a single model and more like a committee.
That feeling is dangerous. A persona is not a control. It is a prompt that changes the distribution of text the model is likely to produce. Sometimes that is useful. Often it is cosmetic. In governance settings, cosmetic diversity is worse than no diversity because it creates the appearance of challenge without proving that any new evidence was inspected.
The recent research points in the same direction. A 2025 study on expert personas found no consistent factual accuracy benefit from telling models to act as domain experts on difficult benchmark questions. A 2026 PRISM paper found the trade-off more explicitly: expert personas can improve alignment and style, but can damage accuracy. Another 2026 study found that demographic persona cues degraded agent task performance by up to 26.2 percent. In multi-agent strategic games, role identities shifted outcomes by up to 90 percentage points even when payoff information was visible. These are not neutral decorations. They are interventions.
That does not mean role assignment is useless. It means the useful unit is not the character. The useful unit is the loss function.
An agent told to “act as a regulator” may produce regulatory-sounding prose. An agent told to “identify claims that would fail supervisory evidence review, using only the facts provided, and return the missing evidence” has a job. An agent told to “be a cautious compliance officer” may become vague and over-conservative. An agent told to “find false de-escalation risk and state which fact would change the route” has a measurable purpose.
The difference is whether the role can be deleted and replaced by a sentence beginning with “protect against.” Protect against unsupported factual claims. Protect against irreversible actions without a recovery path. Protect against hidden data access. Protect against approval drift. Protect against tool failure being smoothed into fluent prose. If the role cannot be stated that way, it is probably theatre.
This is where multi-agent review can still earn its keep. A security lens, a product lens, an assurance lens, and an operations lens really can find different things. But they should not vote on the answer. They should produce evidence questions. What fact is missing? Which claim lacks support? Which control would have to exist? Which action is irreversible? Which assumption is being inherited from a previous case without proof that the cases are actually the same?
The output contract matters more than the persona label. A useful reviewing agent should return a challenge, the evidence needed to resolve it, the part of the decision it affects, and whether the challenge is a blocker or just a question. That is boring compared with a simulated panel discussion. It is also much easier to govern.
The newer multi-agent papers support this boring version. Dynamic role assignment works better when roles are allocated by capability and task fit, not by a fixed cast of characters. Role clarity matters because agents overstep; they drift into one another’s jobs unless the boundary is measurable. Joint prompt optimization matters because each local agent objective can make sense while the whole system gets worse. The system goal is not for each agent to sound plausible. The system goal is for the final decision to become more defensible.
This gives a simple design test for any persona-based agent architecture. What does this role protect against? What evidence is it allowed to use? What output schema forces it to stay in lane? How is its challenge verified? What happens if its recommendation conflicts with the source record? If those questions do not have precise answers, the persona is not part of the control environment. It is part of the theatre around the control environment.
The best use of personas is therefore not to simulate people. It is to force different questions onto the same case. Once the questions are clear, the personality can disappear. What remains is the thing governance can actually use: a bounded lens, a missing-evidence signal, and a decision that still belongs to the accountable owner.
A persona is not a control. At best, it is a lens. The control is what stops the lens from pretending to be the decision.
Sources: Savir Basil et al., Prompting Science Report 4: Playing Pretend: Expert Personas Don’t Improve Factual Accuracy; Zizhao Hu, Mohammad Rostami, and Jesse Thomason, Expert Personas Improve LLM Alignment but Damage Accuracy; Linbo Cao, Lihao Sun, and Yang Yue, From Biased Chatbots to Biased Agents; Viswonathan Manoranjan and Snehalkumar Gaikwad, When Identity Overrides Incentives; Miao Zhang et al., Dynamic Role Assignment for Multi-Agent Debate; Guoling Zhou et al., Improving Role Consistency in Multi-Agent Collaboration via Quantitative Role Clarity; Zhexuan Wang et al., MASPO.