Agents & Action
Actions & rules
The agent is certain: reroute the Paris shipment. It is 2am and no one is awake. Should it just do that, on its own authority? Actions are the governed last mile.
It is 2am. A transit hub outside Paris just closed, and the agent has already figured it out: shipment PAR-7 is about to miss its deadline, there is a clean alternate route, and the fix is to reroute it now. The agent is certain. It is not guessing.
So it should just do that. Right?
Notice why you paused. The agent is right, but something in you still does not want it moving freight at 2am with no one's approval. Rerouting one shipment, fine, maybe. But the same agent, just as certain, might next want to pause dispatch for the whole EU zone. Same confidence. A very different outcome if it is wrong.
That hesitation is the whole concept. An agent that can act and an organization that needs to stay in control are not enemies. One thing lets you say yes to the agent without fearing what it might do, and it is not "trust the model."
Most of the loop is safe. The last box is not.
An agent senses, then it decides, then it acts. Two of those three boxes cannot hurt you. Reading the graph changes nothing. Working out a fix changes nothing. You can let an agent sense and decide all day, and the worst case is a wrong idea sitting in memory.
The last box is different. Act is where a conclusion leaves the model and enters the real world. A shipment actually moves. A ticket actually opens. Money, freight, and someone's whole day all depend on this one box. So this box is the one, and really the only one, that needs a gate.
This is why "can the agent act" is the wrong question. The agent can. The real question is on what terms, and the terms live outside the model, in rules you write.
A rule is not one switch
The easy design is a single switch: autonomous on, autonomous off. It is also wrong, because it treats every action the same way. Off, and your agent can only suggest things, never do them. On, and it can pause the EU zone at 2am.
Good governance sets levels of autonomy instead. The same recommendation meets a rule, and the rule sorts it into one of three lanes. Some things are never allowed. Some things are small enough to run on their own. Some things are allowed, but only after a person approves them. One recommendation, three possible terms, set by policy and not by the model's confidence.
Notice what the rule looks at. Not "is the agent sure." The agent is always sure. It looks at blast radius (how many things the action would affect), at cost, at which action this is, at the time of day. These are facts about the action, checked against a line you set in advance. Confidence is the agent's job. Consequence is yours.
Set the line yourself and watch the same list of recommendations change their terms.
Set the rule
One rule governs every recommendation: auto-fire when the blast radius is at or below the line, otherwise wait for a human. Drag the line and watch the same recommendations change their terms.
Notify the Paris warehouse owner of the delay
blast radius 0 · fired on its own
Reroute PAR-7 around the closed hub
blast radius 1 · fired on its own
Reroute the whole Paris shipment
blast radius 3 · waiting for approval
Cancel and re-book 40 stalled orders
blast radius 6 · waiting for approval
Pause dispatch for the entire EU zone
blast radius 9 · waiting for approval
Drag the line up and more of them run on their own. Drag it down and more of them stop and wait for your click. Nothing about the recommendations changed. The reroute is the same reroute. What changed is the rule. And the rule is the whole difference between an action that happens on its own and one that waits for a signature.
Check yourself
A support agent can give refunds. The team sets it to auto-approve refunds under $50 and send anything higher to a human. A $2,000 refund comes in and the agent is fully sure it is valid. What happens, and why?
Decision to governed action
Everything before this point in the course built toward an answer. The ontology, the knowledge graph, the traversals, the data agent reasoning over all of it. That is the decide box. Actions are the step that lets the decision do something.
In Microsoft Fabric IQ, a data agent (Fabric's name for an AI agent grounded in your data) reasons over the knowledge graph and can propose a fix. From there its recommendation can lead to a governed action: creating a work item, sending a notification, or triggering a flow. The word that matters is governed. Between the proposal and the real world sits a layer of rules and approvals. So the action stays within policy and, where it should, keeps a human in the loop.
Fabric IQ is in preview as of July 2026, so treat the exact list of actions as an example, not a fixed set. The lasting shape is what to remember: a recommendation grounded in the graph, then a rule and approval layer that decides the terms, and only then something real happening. Grounding gives the agent the right answer. The rules decide what it is allowed to do with it.
The one trap
Do not treat confidence as permission. The most dangerous agent is not the one that is often wrong. It is the one that is usually right, because being usually right is exactly what tempts a team to remove the gate. The reroute worked last week, so let it just run. Then one night the same confident agent pauses a zone it should not have, and there was no rule in place to stop it.
The rule is not there for when the agent is wrong. It is there for when the agent is right about the wrong thing, at the wrong scale, at 2am. Keep the gate even when, especially when, the agent has earned your trust.
So you have a gate, and a rule that decides who has to approve. You can almost relax. Almost, because the gate hides one more question no one asked out loud: when the approver clicks yes, whose authority actually acts? Next: who can see it, and who acts.
Do it yourself
Build this step in the interactive Ontology Lab.
Open the lab →Milestone
Finished this concept? Mark it learned to track your progress.