When Your AI Agents Start Fighting Each Other

For two years the AI agent question was "should we have one?" That question is settled. A mid-October 2026 study from OutSystems, run with CIO Dive and KPMG, found that 86% of B2B software buyers purchased AI agent software or services in the past year, a 25% jump. Agents are in production, booking meetings, triaging tickets, reconciling invoices, chasing payments.

Here is the part nobody scoped for. The same research found that only 26% of enterprises feel well prepared to manage conflicts between those agents. Not conflicts with humans, conflicts with each other. When you had one agent, it either worked or it didn't. When you have nine, they start to collide: two of them edit the same record, one reverses what another just did, a sales agent promises a discount a finance agent is built to reject. The failure mode of 2026 is not a dumb agent. It is a room full of capable agents with no referee.

What agent-on-agent conflict actually looks like

This is not theoretical, and it is not science fiction. It is mundane, which is exactly why it slips past planning. The common shapes:

• Double action. A support agent issues a refund. A separate billing agent, reading the same ticket a minute later, issues it again. The customer is thrilled. Finance is not.
• Contradictory decisions. A pricing agent approves a quote to win the deal. A margin-protection agent flags the same quote and claws it back. The customer watched the number change twice in an afternoon.
• The write race. Two agents update the same CRM record or inventory count in the same window. One write silently overwrites the other and the number is now wrong, with no error anywhere.
• Goal collision. A retention agent is told to reduce churn and offers a save discount. A revenue agent is told to protect average contract value and blocks it. Both did their job. The company has no idea which one should win.
• Loops. Agent A opens a task for Agent B, which opens a task back for Agent A. Nobody wrote that loop on purpose. It emerges because nobody owned the boundary between them.

Why this snuck up on everyone

The adoption curve ran ahead of the operating model. Teams bought agents one department at a time: support got theirs, sales got theirs, finance got theirs. Each was scoped, demoed, and shipped on its own. No single person was looking at the whole floor.

This is the next chapter of a problem we have written about before, AI agent sprawl. Sprawl is too many agents. Conflict is what those agents do to each other once they share systems. You can have a tidy, deliberate set of agents and still have conflict, because the question "what happens when two of these touch the same thing" was never anyone's job to answer.

The reason only 26% feel ready is that readiness here is not a tool you buy. It is a set of decisions about precedence, ownership, and arbitration that most companies have not made, because until this year they did not have enough agents in production to need them.

The four controls that actually help

You do not fix this with a smarter model. You fix it with boundaries. Four controls do most of the work.

1. A precedence order, written down. When two agents disagree, which one wins? Finance usually beats sales on money. Compliance beats everyone. Decide the hierarchy before the collision, not during the incident. An undocumented precedence rule is just an argument waiting to happen.
2. Clear write ownership per system of record. For each critical record (a customer, an order, an invoice), exactly one agent is allowed to write, and the rest request changes through it. This single rule kills the write race, which is the most dangerous conflict because it fails silently.
3. A supervisor layer with the final say. A coordinating agent, or a human, that sees the actions of all the others and can hold, reverse, or escalate. This is the practical version of the guardian-agent pattern: AI supervising AI, with a person on the serious calls.
4. Shared, observable state. Every agent writes what it did to one log the others and a human can read. Most conflicts become harmless the moment agents can see each other acting instead of operating blind in parallel.

Five questions before you add agent number ten

Adding agents is easy, which is the trap. Before the next one goes live, make the project answer these:

• What systems will it write to, and who else writes there? If the answer includes a record another agent owns, you have a conflict to resolve before launch, not after.
• What does it do when another agent already acted? Check-then-act beats act-and-hope. An agent that re-reads state before committing avoids most double actions.
• Who wins when it disagrees with an existing agent? A named precedence rule, not "we'll sort it out."
• Can a supervisor see and reverse what it does? If its actions are invisible until a customer complains, you are flying blind.
• Who owns the boundary between this agent and its neighbours? A person, not a diagram. Boundaries without owners rot.

The companies in the 26% are not the ones with the best models. They are the ones who decided precedence, ownership, and arbitration before the agents needed them. This is governance work, the same unglamorous wiring that separates the AI projects that pay back from the ones that stall. It is dull, it is cheap relative to the alternative, and it is entirely in your control.

If you are past your first agent and starting to feel the collisions, tell us how your agents are set up and we will help you draw the boundaries before they draw their own.