World model is the most overloaded phrase in agent design. We use it to mean something specific. A world model is a structured representation of how a particular company makes a particular class of decision.
Not a knowledge base. A knowledge base is documents. A world model is decision logic with named owners.
Not a prompt template. A prompt is text. A world model is data plus policy with versioning.
Five examples, with the company specifics changed but the shape preserved.
01. Pricing exception authority. Sales reps quote inside an auto-approval band of plus-or-minus ten percent against the price book. Outside the band routes to the regional VP. Outside twenty percent routes to the CFO. The world model encodes the band, the routing, the named approvers, and the exception log retention policy. The /finance agent reads this when asked about exceptions.
02. Festival creative cadence. The marketing team posts daily during nine-day festivals, themed by the day. The hub holds the calendar, the theme rules, the historical engagement data, and the brand kit. A Q&A agent answers what to post. A Tool agent generates the asset. The model is what makes both work.
03. Vendor renewal re-review trigger. Any vendor amendment that raises the liability cap above two times annual revenue triggers re-review by legal. Indemnity changes always trigger. The world model encodes the trigger conditions, the routing to the legal lead, and the SLA on response.
04. Slow-mover stock alert. Inventory rotation per SKU has a baseline. SKUs that drop below baseline for fourteen days trigger a clearance review. The world model encodes the baseline, the duration, and the named owner who decides on the discount.
05. Anchor-customer health monitoring. The /delivery agent watches anchor signals weekly. Scope drift, blocked decisions, executive sponsor engagement. Each signal has a threshold and an escalation owner. The world model encodes which signals matter, what the thresholds are, and who owns each escalation.
Across all five, the pattern is the same. Decisions a company already makes. Encoded as data and policy. Owned by named operators. Read by role-scoped agents. Versioned in a repo, not a prompt.
This is the work. Most of the deployment is here, not in model selection.