Advanced · Research

World-Model Agents

Visual diagram · Math · (in preparation) · Worked example · 3 difficulty levels.

TL;DR. Agents that plan by simulating outcomes inside a learned world model before acting in the real environment.

Technical Definition

Agents that plan by simulating outcomes inside a learned world model before acting in the real environment.

How it works

A world-model agent keeps an internal, learned simulator of its environment — how a codebase behaves under edits, how a robot's gripper responds, how a market reacts — and rolls out candidate plans inside the simulator before committing to one in reality. This converts trial-and-error into imagine-then-act: fewer real-world mistakes, better handling of irreversible actions, and the ability to train on imagined experience. The 2026 wave combined video-generation world models with LLM planners. The central risk is model-reality gap: a planner that over-trusts a flawed simulator optimizes for a world that does not exist, so deployments bound the horizon of simulated rollouts and verify early steps against reality.

Visual Explanation (cycle)

Observe → update world model → simulate candidate plans → score outcomes → act on best plan → observe

Related Concepts

• Simulation — Synthetic environments that let agents or models train and be evaluated without real-world risk.