--- title: "Two agents disagree: how a swarm settles a dispute" url: "https://kyberagent.com/blog/two-agents-disagree" description: "Two sub-agents return contradictory results and the planner must pick one. Why a vote is the wrong tool, what evidence on a claim means, and when the question goes to you." --- # Two agents disagree: how a swarm settles a dispute October 22, 2026·4 min read·Editorial team Two rounded compute cells facing each other across a small balance scale, a third cell with a magnifier lens standing behind it A swarm is checking a batch of supplier invoices against the orders they claim to fulfil. Two sub-agents handle neighbouring parts of the batch. The first reports that a certain delivery was short by one pallet and the invoice should be disputed. The second, working from the warehouse log, reports that the same delivery arrived complete and the invoice is fine. Both reports are fluent. Both cite something. Both are confident. The planner has one slot on its list for that invoice and two incompatible answers to put in it. In a lot of agent systems what happens next is that the planner reads both, writes a sentence that sounds like a decision, and moves on. Afterwards nobody can say which report won, or why. ## The real cause: a claim without its evidence is just a voice A sub-agent’s report is a conclusion. If it comes back as prose alone, the planner has nothing to compare except the prose, and prose from two instances of the same model is going to look equally plausible. The planner then does what people do with two confident voices: it picks the one that sounds more sure, or the one it read last, or it asks a third instance and treats the majority as truth. None of those is a rule. They are habits, and a dispute settled by habit cannot be checked later. The fix is not a smarter planner. It is reports that carry what they are built on. ## How disputes go wrong **The vote.** Ask three agents, take two out of three. It feels fair and it is nothing of the kind. Three instances of one model share one set of blind spots, so a vote measures how the model leans, not what is true. It also costs three times the work to produce an answer that is no better supported than one. **The confident one wins.** A report that says “clearly” and “confirmed” beats a report that says “the log suggests”. The hedged report was the careful one. Rewarding tone trains the swarm to stop hedging, which is exactly the wrong lesson. **The planner merges them.** Faced with “short by one pallet” and “arrived complete”, the planner writes “delivery mostly complete, minor discrepancy possible”. That is not a reconciliation. It is a third claim with no evidence of its own, and it hides that a dispute ever happened. **Last report in.** The planner’s context holds the first report when the second arrives, and the second is fresher. Order of arrival becomes the deciding factor without anyone choosing it. **The dispute disappears.** Whatever the planner decides, the losing report is dropped and the plan shows one tick. A month later the invoice turns out to be wrong, and there is no trace that one agent said so at the time. ## What a swarm should do instead Our design is that a dispute is a case to be decided, with a procedure, rather than a disagreement to be smoothed over. **Every claim carries its evidence.** A sub-agent’s report is not just a conclusion but the material it rests on: which file, which lines, which tool output, obtained in which step. The first agent points at the delivery note. The second points at the warehouse log entry. Now the planner is comparing two documents, not two voices. Reports without evidence attached are returned as incomplete, not weighed. **A verifier with tools, not a third opinion.** When two evidenced claims conflict, the planner does not ask a third agent what it thinks. It starts a verifier sub-agent with a narrow brief: here are two claims and their sources, find out which source is wrong. It gets the tools to open both documents, and only those. The verifier’s job is to produce a finding about the evidence, not a vote about the conclusion. **Equal evidence goes up, not sideways.** Sometimes the verifier comes back and says both sources are internally consistent and simply disagree: the delivery note and the warehouse log were filled in by different people. That is not a problem the swarm can solve by spending more compute. The item is marked disputed and goes to KyberBot, which asks you, with both pieces of evidence laid side by side. The swarm waits. It does not guess. **The record shows the dispute.** All of it goes into the [trajectory](https://kyberagent.com/blog/every-step-on-the-record): the two reports, the evidence each attached, the verifier’s brief and its finding, and either the decision rule that settled it or the question that went to you and the answer that came back. The plan shows the item as settled, with a pointer to how. A month later, when the invoice question returns, the losing report is still there, and so is the reason it lost. **Every party in its own cell.** The two sub-agents and the verifier each run in their own [KyberCell](https://kyberagent.com/blog/kybercell-explained). The verifier cannot be swayed by the first agent’s scratch work because it never sees it. It sees the evidence that was formally attached, which is the point. The limit, said plainly: this is a design we are building, not a released feature, and it does not make the swarm right. A verifier can be wrong about which source to trust. What the procedure guarantees is narrower and more useful: a dispute is never settled by tone, by order, or by silence, and you can always see how it was settled. ## A test you can run today Give an agent you already use the same factual question twice, phrased slightly differently, in two separate sessions. Pick a question whose answer lives in your own documents, where two sources might reasonably disagree: a date, a quantity, a name. If the two answers agree, try again until they do not; on real documents it will not take long. Then ask a third session to resolve the two answers. Look at what comes back. If it says which of the two is right without pointing to the source that settles it, you have a vote dressed as a verdict. If it points at a source, you have the beginning of a rule, and the rest of the rule is deciding what happens when the sources are equal. That decision, and the budget that bounds how much a swarm may spend before it makes it, are the subject of [when a swarm should stop](https://kyberagent.com/blog/when-a-swarm-should-stop). How the planner, the cells and KyberBot fit together is on the [KyberAgent homepage](https://kyberagent.com/). - sub-agents - verification - evidence - escalation - trajectory ## Keep reading - [When a swarm should stop: time and token budgets for a task](https://kyberagent.com/blog/when-a-swarm-should-stop) - [Agents on Android XR: background tasks and the cell](https://kyberagent.com/blog/agents-on-android-xr)