Illustrative scenario. The sales agent confirms a 15% discount and 48-hour delivery to a customer. A day later the support agent tells the same customer delivery takes seven days and that discount doesn't exist. The customer forwards both replies with one line: "Which one is right?" Nobody at the company can say without opening three screens.
The thesis in one line: when two agents contradict each other, the model is almost never the problem. They read different truths and nobody owned the decision. You fix it with one source per fact, a precedence rule and a human referee for what the rule doesn't cover, not with a third agent that has an opinion.
Conflicting AI agents in the company: is it the model or the source of truth?
The first move is to blame the model and try another one, or tweak the prompt. It rarely works, because both agents answered correctly from what was in front of them: sales read the offer in the CRM; support read the policy in the knowledge base and the lead time in the order system. Three copies of reality, each right in its own place and none in sync with the others.
This isn't an exotic AI problem. It's what happens in any company with two departments and two spreadsheets. The difference is that an agent answers in seconds and at scale, so the contradiction reaches the customer before anyone sees it. Cognition, the company behind Devin, makes the point for multi-agent systems in general: if each agent decides with its own isolated context, their decisions can collide ("Don't Build Multi-Agents," Cognition). If you're still unsure you need several agents, start with one agent or many?; here we assume you already have them.
And it's expensive for a reason that isn't technical. A contradiction that reaches the customer is a written promise from the company and proof that nobody is in charge of the answers. Fixing it once costs a call from someone with judgment; repeated, it costs you the customer's trust in the automated channel, and they call every time.
Why isn't a supervisor agent enough?
It's the fashionable answer: an agent on top that compares the replies and decides. It does detect contradictions, and that's useful. But deciding who is right requires knowing what's true, and if the supervisor reads the same three copies, it's a third voice with an opinion. With luck it gets it right; without luck you have the same conflict one level higher. A supervisor is a good detector and a bad judge.
How do you resolve a conflict between agents without the customer refereeing it?
- Inventory the facts both agents can state. Price, discount, lead time, stock, return terms. The facts two agents can say about the same customer are the ones that contradict; start there.
- One owner system per fact. Lead time lives in orders; the current discount, in the price list; the policy, in a single document. Every other agent reads from there, not from its own copy. The connection is designed in how to connect AI to your CRM, and what each agent reads and writes in what permissions to give an AI agent.
- A written precedence rule. If two sources differ, operations wins on lead times and finance wins on prices. This isn't a conversation between agents: it's a line in a document anyone can read.
- Agents write to the record, not to each other. If sales promises something, it lands in the customer record where support sees it. Every message from one agent to another loses context; what's already in the record isn't lost. Shared memory has its own design: an AI agent's memory.
- Whatever the rule doesn't cover goes up to a person, with both versions in front of them. A named referee decides in minutes if they see both answers and their sources; without them, they have to rebuild everything.
That last step is the right level of autonomy for contradictions: faced with two truths, the agent doesn't choose, it flags. It's the same principle as what an agent does when it doesn't know, applied to two agents instead of one.
| Type of conflict | Typical cause | Who decides |
|---|---|---|
| Different fact (lead time, price) | Each agent reads its own copy | The single source; fix the origin |
| A promise the other can't keep | Sales commits without seeing capacity | The precedence rule: operations validates lead times |
| Ambiguous policy | Two documents say different things | A person; then one document is retired |
| Two versions already sent to the customer | The conflict reached the customer | A human contacts them and sets the right one |
What do you measure to know the conflict is under control?
Three things, none of which needs a market figure: how many contradictions are detected per week, how many get resolved by the rule without bothering anyone, and how long a human takes to decide the ones that escalate. If the first falls and the second rises, the system is learning; if they all rise together, you're adding agents faster than you're organizing the data. Who reviews those numbers and who fixes the origin is a maintenance question: who maintains AI agents.
When don't you need a referee?
When the two agents share no data and no customers, there's nothing to contradict; and when the error is cheap and reversible, like a different product recommendation, a precedence rule is enough. The human referee is for what commits money, deadlines or the company's word.
And if you already have several agents with no clear division of work, the problem is the system: orchestrating several AI agents is the job of giving each one its place, its source and its rule. For the ones that touch the customer, the AI sales agent and the AI support agent should read from the same record from day one.
The line to take with you: two agents that contradict each other don't need a boss, they need one truth. Give it to them once, in writing, and keep a person for the day the truth isn't enough.
Sources (accessed October 26, 2026): "Don't Build Multi-Agents," Cognition. The opening scenario is illustrative, not a real case.