Almost every multi-agent AI system is built to do one thing with disagreement: get rid of it. Vote it down. Debate it out. Average it away. Treat the model that dissents as a fault to be tolerated, the way a distributed system tolerates a bad node.
A new paper out of Warsaw argues that for the questions actually worth putting to a council, that goal is backwards.
The wrong number to optimize
The paper is arXiv:2606.04223, "Consensus is Strategically Insufficient," posted June 2 and accepted to the LAMAS&SR workshop at FLoC 2026. Michał Wawer and Jarosław A. Chudziak, of the Laboratory of The New Ethos at Warsaw University of Technology, make a narrow, sharp claim: the dominant objective of multi-agent systems — drive agents toward agreement through voting, consensus protocols, debate, or fault-tolerant aggregation — is the wrong objective for value-laden tasks.
Their reasoning is direct. When two models split on whether a post crosses a moderation line, or whether a borderline claim is true enough to publish, the split is often genuine normative uncertainty, not one model being broken. Collapse it into a majority vote and you have thrown away the most informative thing the council produced. You get a clean answer and lose the reason it was hard.
This is not a complaint about accuracy. It is a complaint about what counts as the output.
Stop reading the verdict. Read the shape.
The paper's move is to stop treating the final answer as the product of deliberation. The product is the structure of the disagreement itself.
Consensus tells you the models landed in the same place. It does not tell you whether they took the same road.
Wawer and Chudziak build a knowledge-representation layer that abstracts each agent's reasoning trace and its binary decision along two axes: how similar the reasoning is, and whether the conclusions agree. Crossing those two axes gives four symbolic states — and each one means something different about the question you asked.
Convergent agreement — similar reasoning, same conclusion. The models walked the same path and ended in the same place. This is the case everyone is optimizing for, and it is the one that needs the least attention. Commit and move on.
Divergent agreement — different reasoning, same conclusion. The models agree on the answer but got there by unrelated routes. This is arguably the strongest signal a council can produce: independent roads converging on one destination is harder to dismiss than a chorus that all reasoned the same way.
Convergent disagreement — similar reasoning, opposite conclusions. The dangerous one. The models reasoned almost identically and still split at the very end. That is a tell that the decision turns on a threshold or a value judgment that no additional reasoning will resolve. More debate will not fix it, because the models already agree on everything except the line.
Divergent disagreement — different reasoning, different conclusion. Genuine pluralism. The models are looking at the problem through different frames and reaching different ends. This is the case a vote flattens most violently, and the one most likely to deserve a human.
Four states, four different things to do next. The paper does not leave them as a taxonomy: the states drive defeasible strategic routing rules, instantiated in a content-moderation setting, that decide whether the system commits, flags, or escalates. The shape of the disagreement determines the handling.
This is the Chairperson's actual job
Strip the formalism and this is a claim Shingikai has been built around: the synthesis layer's job is not to produce one blended answer. It is to decide what to do with the disagreement.
A council that returns "the answer" has discarded the diagnosis. A council that returns "the models split the same way every time on this question" has told you something about the question itself — that it is a value call, not a fact, and that no amount of additional argument is going to settle it.
The four-state map is a precise vocabulary for what a synthesis surface should show a user instead of a single confident paragraph. "Two models disagreed" is almost useless. "Two models reasoned the same way and still split" tells you the call is yours to make and tells you why.
What the gap looks like in practice
Ask one model whether a borderline post violates policy and you get a yes or a no. Run a council and look only at the vote, and you get a tally — three to two, decision rendered, reasons gone.
Run a council and look at the state, and you learn something the tally hides. If the models reasoned identically and still split, that is convergent disagreement: the call is a threshold choice, not a factual one, and it belongs to a person who can own where the line goes. If they reasoned differently and still agreed, that is divergent agreement, and you can act on it with more confidence than any single model would have earned.
Same five models. Same question. The difference is entirely in whether the system preserved the shape of the disagreement or flattened it into a number.
The disagreement was always the reading
The paper bridges two worlds that usually ignore each other — the sub-symbolic reasoning of LLM agents and the symbolic structure of classical knowledge representation — and the bridge is the part worth borrowing. Most councils still treat divergence as a problem to be minimized on the way to a clean output. Wawer and Chudziak treat it as the output.
That is the reframe. A council is not a machine for manufacturing agreement. It is an instrument for measuring the kind of disagreement you are dealing with, so you know whether to trust the answer, argue more, or stop and decide for yourself.
The disagreement was never the noise. It was the reading.
If you want to watch models split the same way on a question and see the shape of it instead of a flattened vote, that is what a council is for. Try it free, no signup — shingik.ai.