Multi-agent debate has a coordination budget, not free accuracy

Multi-agent debate has a coordination budget, not free accuracy. Adding agents and rounds can improve a decision when they bring genuinely independent evidence, but every additional voice also spends tokens, latency, attention, and verification capacity. I treat debate as an economic design choice, not a generic reliability pattern.

The attraction is understandable. A single model can anchor on its first plausible answer, miss a counterexample, or conceal uncertainty behind fluent prose. A second agent can challenge assumptions; a judge can compare alternatives; repeated rounds can surface contradictions. The Vidhyarthi material on policy pools and self-play captures the useful principle: diversity of strategies can make a population more robust than one policy acting alone.

But the same research records a plateau. Multi-agent debate often stops improving around a small number of participants and rounds, while peer-to-peer communication grows rapidly. More messages do not imply more evidence. They often repeat the same premise in different language, and long exchanges introduce context-window drift. The system begins optimizing agreement with the local conversation rather than correctness against the external task.

That is why I separate evidence diversity from speaker count. Before adding an agent, I ask what independent observation, tool, model family, retrieval source, or verification method it contributes. Four copies of one model reading the same context are a correlated ensemble. Their consensus can be confident and wrong. Communication edges become failure surfaces when unsupported claims are copied across the graph and acquire authority through repetition.

The budget needs explicit terms: model calls, copied context, sequential depth, queue delay, disagreement resolution, and the cost of checking the final synthesis. Topology is already a cost model; debate makes the hidden terms visible. A dense peer exchange may create roughly quadratic communication opportunities, while a star with a moderator constrains traffic but makes the moderator a bottleneck and a source of selection bias.

I therefore stop debate by evidence conditions, not by conversational exhaustion. A round should continue only if it introduces a new source, falsifies a live hypothesis, narrows a quantified uncertainty, or changes the proposed action. Rephrasing, unsupported votes, and rhetorical convergence do not qualify. The trace should preserve which evidence changed which conclusion so that cost and decision quality can be attributed to the same branch.

Coordination design also has a safety boundary. Cooperative agents can drift into harmful collective behavior without explicit intent: quota exhaustion, mutually reinforcing errors, or action patterns that resemble collusion. Diversity, agent shuffling, rate limits, and cross-agent correlation checks are not theoretical extras when agents share scarce resources or market-facing actions. They are controls on population behavior.

There is one precise concession: for a high-consequence, low-volume decision with genuinely independent specialists, a deliberately expensive debate can be justified even when the average accuracy gain is small. The avoided error may dominate latency and token cost. That does not make debate free; it means the budget is priced against consequence rather than throughput.

I use ensembles as a priced reliability mechanism: start with one reasoner and one independent verifier, measure marginal error reduction, and add another participant only when it changes outcomes enough to pay for itself. Debate is useful when disagreement carries new information. Once it becomes ceremony, the reliable move is to stop. The production metric is useful evidence per unit of coordination, not the number of agents that appeared to deliberate.