What happens when two AI agents disagree?

Nothing happens automatically unless a rule people wrote settles it. A disagreement between agents is a signal that the facts are unclear or one agent is wrong. Where the agents relied on sources of different strength, the stronger source wins by rule. Otherwise the case is held for a named person, who sees both positions and the facts behind each.

Why do agents disagree?

Usually for one of four reasons.

• Different sources. One agent read the purchase order; another read the supplier's invoice. They are describing different documents.
• Different timing. One agent read a balance before a payment cleared; another after.
• Different roles. A matching agent is looking for reasons to pay; a fraud agent is looking for reasons not to. Both are doing their jobs.
• One is wrong. An agent misread a field, applied a stale rule or was steered by an input.

The first three are information. Only the fourth is a fault. Treating every disagreement as a fault to be suppressed loses the information.

Why not let one agent win, or let a third agent vote?

Letting one agent always win makes the other pointless. If the matching agent overrides the fraud agent, you do not have a fraud check.

A third agent breaking the tie sounds sensible but is weaker than it looks. Agents built on the same model, fed the same inputs, tend to make the same mistakes. Two out of three agreeing may just mean two out of three misread the same thing. Voting works only when the voters are independent, which agents sharing a model and sources are not.

How should disagreements be settled?

1. Record both positions, each with its sources and the agent's certificate.
2. Apply source rules. If one position rests on a fact from your own system of record and the other on an external claim, the rule people set decides. A two-way match between a purchase order and an invoice is a source comparison of exactly this kind; see what a two-way match is.
3. Re-read the facts. If timing caused the disagreement, a fresh read may resolve it.
4. Hold the rest. Anything the rules cannot settle goes to a named person.
5. Act only on the outcome. Neither agent proceeds on its own view while the disagreement is open.

What may the agents do, and what stays with a person?

What does this look like in practice?

Take a wholesaler's accounts payable. The matching agent finds an invoice matches its purchase order and goods receipt within tolerance and proposes it for payment. The fraud agent notices the invoice's bank details differ from the supplier master file and flags it.

The platform records both. The source rule says bank details from the master file outrank details printed on an invoice, which is external. So the invoice can be paid, but only to the master file account. The changed bank details on the invoice go to a person to verify by calling the supplier on a known number. If they turn out to be real, the master file is updated through the approved route. If not, the business has caught a fraud attempt.

Neither agent was wrong. The disagreement was the control working. The wider point is made in name checks are only as good as what happens after a mismatch.

Often agents disagree because the systems they read disagree. That is a reconciliation problem, and agents can help work it, as covered in how to reconcile records between two systems with an AI agent. A related case is when an agent simply does not have the answer, covered in what happens when a support agent does not know the answer.

Where Pryme fits

In Pryme, every fact carries a provenance class, so source rules can settle many disagreements without a person. Where they cannot, the action is held for a named person, and both agents' proposals are filed in the same append-only record. The full design is in the governed agentic operating system guide.