World Models for Agents: Why the Interesting Ones Are Graphs
A language model can describe a kitchen in perfect prose and still fail to plan boiling an egg. The gap is not knowledge, it is a world model: an internal structure that says what actions do, what they require, and what they change. For a decade the frontier of world modelling was pixels — PlaNet, the Dreamer family, MuZero learning to roll out futures in latent space. For agents that act in text and tool environments, the interesting frontier is quieter and older: graphs.
A world model does two jobs
The recent survey Understanding World or Predicting Future? (with its companion reference list) frames the field cleanly. A world model does two things: it maintains an internal representation of the environment's state, and it predicts the future given actions. Most failures trace to confusing the two — a model that predicts plausibly but carries no stable state will drift, and a model that holds state but cannot simulate will not plan.
That distinction matters more as world models move from games into agentic software tasks, where the "physics" is a mix of API contracts, preconditions, and user intent.
Three levels, four sets of laws
Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond proposes a taxonomy that is worth adopting wholesale. It defines three ascending levels:
| Level | Name | Capability |
|---|---|---|
| L1 | Predictor | One-step, local transition prediction |
| L2 | Simulator | Multi-step, action-conditioned rollouts that respect domain laws |
| L3 | Evolver | Revises its own model when predictions fail |
It then crosses these with four law regimes — physical, digital, social, and scientific. A robot needs physical laws; a coding agent needs digital laws (a database write is not idempotent by default); a negotiation agent needs social laws; a research agent needs scientific ones. The taxonomy is a useful lens because most deployed "world models" are L1 predictors dressed as simulators, and L3 — genuinely self-revising models — remains rare.
The latent-failure problem
The cleanest way to see why this matters is a benchmark built to expose it. SIMMER evaluates symbolic world modelling in LLM planning using a kitchen domain: 77 actions, 262 objects, roughly 46,800 interactions. The findings are bracing:
- Frontier LLMs produced at most 17% error-free plans.
- Up to 56% of generated plans contained latent failures — mistakes invisible until execution, with the majority being irreversible.