Guest Post: The Power of Compact World Models (CWMs) for Agentic AI

We’re hearing a lot about Agentic AI these days. If you work in an office and do most of your work on a computer, chances are that someone will soon offer you an AI Agent to help you manage documents or software. But what if your work involves managing things in the real world like scheduling and completing maintenance for a fleet of heavy vehicles? You need an AI Agent that knows something about the world outside a computer. It needs a World Model.

World models are—as a rule—very ambitious. But they don’t have to be. You can start small, with a compact world model (CWM) of what you know about the world of vehicle maintenance—maintenance bays, mechanics, parts, and regulations.

A CWM represents a little slice of the real world that embodies useful knowledge or expertise in a specific, limited domain. Unlike systems modeling and boil-the-ocean approaches, a CWM provides low-hanging-fruit value that you can build on iteratively, one decision at a time, instead of requiring a big engineering effort before you have any idea of its value.  

(Note that we could reasonably call CWMs “Small World Models” (SWMs) or “Small Reasoning Models” (SRMs) instead.)

How to build it

How do you capture what you know? Start by building a simple Causal Decision Diagram (CDD) of the actions you take in maintenance scheduling (put these on the left-hand side of a piece of paper) and the outcomes you expect to achieve (list them on the right). Your outcomes might include cost, likelihood of breakdown, or adherence to safety regulations. Your actions might include choices like scheduling people, facilities, vehicles coming out of service, and parts.

In addition, you’ll need to include those external factors that are partially or totally out of your control like availability of people, parts, and facilities, and changing regulations. People make your job harder when they get sick or quit. Parts become a nightmare when there are supply chain disruptions or costs skyrocket due to tariffs. Your facilities may become unavailable due to power outages caused by unexpected events like fire hazards or hurricanes. Or — to your benefit — federal regulations may loosen giving you new efficiency options.

Keeping it compact and simple

Stick to the actions that are under your control, the outcomes for which you’re responsible, and the external factors that are both relatively easy to measure and which, if they change, would mean you’d take a different action. And this is a situation where the perfect can be the enemy of the “good enough”: often a minimum useful model is a great improvement over keeping it all in your head or in text alone. You can always add more later. Connect up the causal chains with arrows and measurements going from left to right and you have a model of your expertise — a low-fidelity digital twin — when you’re wearing your maintenance scheduler hat.

(Learn more about how to draw decision diagrams in my DI Handbook, or we can help you at Quantellia through solutions and courses).

Running the digital twin simulation

Now you want to know the best actions to take in various possible future situations. So you simulate your model to get “data from the future” (or, to be more precise, “data from many possible futures.”) You’re looking for how optimal scheduling actions will change as the external factors take on different values. Some of these, like 50% tariffs or drastically reduced regulations, you may have never encountered before (and so any data you have based on the past is obsolete). The good news: you’re using a forward in time simulation model, based on the cause-and-effect structure of your situation, so you’re not subject to the irrelevant historical data problem that plagues machine learning and statistical methods in changing circumstances.

Your CWM is robust because you’ve accounted for expected volatility. And if you’ve used a general-purpose Decision Intelligence simulator like Quantellia’s World Modeler™, it’s a resilient model because as your maintenance world changes, you can alert to changes that are big enough that your actions won’t lead to your expected outcomes. You can then add new external factors, new values of existing externals, and/or new actions, and simulate again to get an updated or improved model.

A CWM as the reasoning component of agentic AI

Your maintenance compact world model (CWM) can form the brain of an Agentic AI system, which reasons through your maintenance choices using the model, can answer questions about its reasoning, then takes actions itself — like maintenance scheduling — to make it so.

And now here’s your challenge: this isn’t just about maintenance but thousands of use cases for which AI and data have never been used in the past. What’s something you care about for which you can list the actions, outcomes, and externals? You can use the DI methodology to build CWMs for choosing a training program for your athletes , choosing which parts of your telecom network to build out in which order, picking a price for your product, creating a greenhouse gas emissions policy for your company, deciding which sweet potatoes to pack for your latest grocery store customer, and many many more. The best part: because this approach is compact (and much easier than full world models), you can get a long way with just a bit of effort.