The World Model: How Agents Get Their Bearings

Building an agent is easy now. Building one that stays competent as it scales, without burning tokens like it is trying to set a record, is the hard part. The single biggest lever we have found for that is not a sharper prompt or a longer list of rules. It is giving the agent a world model.

Here is what we mean, what it did, and where we are still guessing. Take it as one team's field notes, not a law.

What a world model is

A world model is a model of the world the agent works in, kept in the code and handed to the agent before it acts.

Not instructions. Not "do this, avoid that." A description of reality. Who the company is and what we are building. Who a given client is and what their work actually looks like. What each object in the database means. How the platform is wired. What the other agents around it do. How we expect the people on the other end to be treated.

Each piece exists at two depths. A surface line for orientation, and a deep dive for when a task actually needs the detail. The agent gets the surface by default and reaches for depth only when it has to.

Why we reached for it

We had been improving agents the usual way. A miss produced a new rule. A new client produced their own rules. The prompt kept growing and the agents stayed stuck at okay. More rules did not seem to buy more reliability. If anything, they tracked with more drift.

So we stopped writing rules and started describing the world instead.

What it did to performance

We seeded a range of agents with the world model, agents that had nothing to do with each other, and output quality went up around the same time. Not one agent we had been babysitting. Many of them, at once. Behaviours that had been inconsistent for weeks settled down. And the instruction half of the prompt, the part that had been swelling for months, shrank to a fraction of its old size.

Honesty about the evidence: we did not run a clean A/B with everything else held constant. This shipped alongside other changes, and "quality went up" is partly judgment, not purely a metric. So treat the effect as a strong impression, not a proven number.

Why we think it works

Our best guess is that the agents were never short on rules. They were short on a model of where they were.

“

A rule tells an agent what to do in one case. A world model lets it work out the right thing in the cases you never wrote a rule for.

Give an agent the client, the objects, and the goal, and a lot of the rules we used to write turn into things it can infer. That would explain both results at once. Quality rose because the agent could reason from context instead of pattern matching a rulebook, and the prompt shrank because we stopped spelling out every case. That is a hypothesis. It fits what we saw. We are not claiming it is the only reading.

Why the concept scales

Competent and efficient at the same time is the hard trade, and a world model is how we are trying to win both. It is easy to make an agent competent if you are willing to be wildly inefficient. Fill the context window, brute force every case, pay for the tokens. It is just as easy to make one cheap and useless. A world model splits the difference on purpose. Surface snippets keep every agent oriented for almost nothing. Deep dives load only when a task earns them. The agent reasons from a compact picture of reality instead of rereading a rulebook it half understands.

It also compounds. One world model serves every agent you run. Describe a client or a database object once, and each agent inherits it. New agents start oriented instead of starting from a blank prompt.

If you try one thing

Before you write your agent's next rule, ask whether it actually understands the world it is working in. For us the answer was usually no, and fixing that did more than any instruction we ever added.

We build credentialing agents that work across payers, provider types, and clients who each define "done" a little differently. A rulebook for all of that never ends. A world model might. That is the bet we are making, and so far it is holding.