A scheduling agent wants to keep a production line running. A maintenance agent wants to shut it down. A quality agent wants to hold the product coming off it. All three could have a valid reason. All three could be working exactly as designed. And you still have a decision to make. This is one of the more interesting challenges in building AI into operations. Once you move beyond asking a chatbot questions and start giving agents access to systems, tools, and workflows, you have to decide how their decisions fit together. Giving each agent a job is relatively straightforward. Defining what happens when those jobs collide takes considerably more thought. Consider a hypothetical production scenario. An order is due tomorrow, and the schedule has very little room to move. Equipment readings suggest a developing problem, while recent quality results show a trend that needs investigation. The scheduling agent recommends finishing the order before stopping. The maintenance agent recommends an inspection now. The quality agent recommends holding the affected production until someone verifies the results. You could ask another AI agent to review the recommendations and choose a path. But that immediately raises another question: what authority does that agent have, and what rules is it applying? Putting another model above the first three doesn’t resolve a missing business decision. Start with what each agent is allowed to do When I think about agents in an operational environment, I want a clear distinction between what they can observe, what they can recommend, and what they can change. A scheduling agent might be allowed to generate an alternative production plan without approval. Moving a committed order, changing a material allocation, or publishing that schedule could require different permissions. A quality agent might be allowed to flag a suspected issue, while placing or releasing a formal hold follows a defined workflow. Those are separate actions with different consequences. That boundary needs to exist in the application and its tools. If an agent is only authorized to propose a schedule, the system should prevent it from publishing one. A sentence in a prompt asking it to seek approval isn’t enough control for that job. The same applies to the information it can access. An agent needs the context required for its task, with permissions that match its role. Giving every agent access to everything makes it harder to understand and control how decisions get made. They also need to agree on what the situation actually is Before deciding which recommendation should take priority, I’d want to know whether the agents are looking at the same version of events. Is the scheduling agent using the latest equipment status? Does the maintenance agent know the next planned shutdown is two hours away? Is the quality agent evaluating current production, or results from a previous run? A disagreement can come from different priorities. It can also come from stale data, missing context, or different definitions. “Available” is a good example. A machine can be mechanically available but waiting on sanitation. Material can physically exist in inventory but be allocated to another order or under a quality hold. A schedule that treats either as ready to use may look excellent until someone tries to execute it. For a consequential recommendation, I want the system to retain the underlying records, timestamps, assumptions, and missing information. That gives the next step something concrete to evaluate. A confident paragraph saying “continue production” doesn’t tell me enough. Some decisions belong in business rules There are conditions I would want handled through explicit rules and enforced workflows. If an active quality hold prevents material from being used, the scheduling agent should treat that material as unavailable. It shouldn’t be able to argue its way around the hold because an order is late. Where the business has defined a mandatory stop condition, that condition needs to be enforced by the appropriate control. It shouldn’t depend on which agent produces the most persuasive explanation. Within those boundaries, there may still be several reasonable options. Could we move the order to another asset? Inspect during a changeover? Run a different product while the issue is investigated? That is useful work for agents: gathering context, evaluating alternatives, and showing the consequences of each option. The technical design needs to distinguish those choices from conditions where the decision has already been made through policy, process, or a required control. The handoff is part of the architecture Suppose the maintenance issue requires stopping the asset. What happens to the scheduling recommendation that was generated five minutes earlier? It needs to be invalidated or reconsidered. Otherwise, you have one part of the system acting on a decision another part has already made obsolete. That means thinking about how agents share state, how changes trigger new work, and how the system checks that a recommendation is still valid before carrying it out. There’s also the problem of two agents trying to change the same thing. If both submit an update to a production order, the application needs a defined way to handle that conflict. Letting the last update win could quietly erase an important decision. These are familiar software engineering problems, but adding agents gives you more paths through them. The model may choose different tools or propose a different sequence of actions. The application still has to protect the integrity of the underlying records. “Ask a human” needs more design than it gets There will be situations where someone needs to make the call. But an escalation that simply says “agents disagree” has handed the investigation back to the person. A useful escalation should explain the conflict, show the evidence, identify the available options, and make the required decision clear. It also needs an owner. Who receives it? How quickly do they need to respond? What can safely continue while they’re deciding? What happens if they aren’t available? Those questions belong in the workflow design. They shouldn’t first come up when production is waiting. And once the person decides, that decision needs to flow back through the system so the agents, records, and pending actions reflect it. Test the collision Testing each agent individually only tells you part of the story. I’d also want to test what happens when the quality hold arrives after a schedule has been proposed but before it is published. When equipment status changes halfway through a workflow. When one agent loses access to a tool. When an approval expires, or an action succeeds but the confirmation fails. Those scenarios reveal whether the system can recover without duplicating actions, using stale recommendations, or leaving people unsure of what actually happened. A convincing demonstration might show three agents completing three tasks. An operational system needs to handle the moment those tasks interfere with one another. That’s where this gets interesting to me. The disagreement itself can be useful. It can expose a tradeoff that deserves attention. But the software needs a clear way to turn that conflict into a controlled, traceable decision. Before you give AI agents the keys, decide who has the right of way. This post first appeared on Tracy’s blog.