World Models · 2 June 2026 · 14 min read
Agent Capabilities Are Growing, But The Architecture Hasn't Caught Up.
Task agents act. Decision agents have to reason first, and that needs a different foundation.
Two-thirds of organisations report measurable productivity improvements from AI, with the majority crediting agents: coding agents ship production software, legal agents review, redline, and summarise contracts, customer support agents resolve tickets at scale. Acceptance is accelerating, and with it so is trust. Agents are being handed decisions of increasing consequence, in domains of increasing complexity and risk.
Capabilities outpacing architecture
This shift is quickly outpacing current task-agent capabilities, signalling a need for an adjustment in architecture. The boundary of what agents are being asked to do is changing categorically, and the infrastructure underneath most deployments has not changed with it.
Two kinds of agents
We all use agents differently and employ different prompting strategies to improve outputs, but there are limits on the agent capabilities themselves, and not all agents are the same. The distinction that matters is between task and decision.
Task agents operate in bounded domains with self-contained outputs: summarisation, review, support, shipping products. When a task agent is wrong, the output is often clear, a bad diff, a missed clause. The mistake is visible, local, and recoverable before it touches anything downstream.
Decision agents are different in kind. A decision agent anticipates, strategises, and commits to a course of action that propagates across work. Consider a supply-chain case: scraping the internet for suppliers and scoring their reliability is a task, while deciding which supplier to order from is a decision. That choice carries downstream ramifications across cost, lead time, relationships, inventory positions, and customer commitments, simultaneously, across every part of the business that touches it.
Decisions require imagining alternatives, reasoning about consequences, and choosing under uncertainty. That is not the same problem as producing a bounded output. When a decision agent is wrong, the business moves in the wrong direction, across margin, inventory, customer commitments, supplier relationships, and cash. The blast radius is the entire operation, and that changes what needs to exist underneath.
Why current frameworks are architecturally incomplete
The dominant agentic frameworks were built for task agents. They are good at acting, but they cannot simulate before they act. For task agents this is fine, the output is checkable. For decision agents, acting without simulating first can lead to major failure and missed value. Only 34% of organisations use AI to deeply transform how they work, fewer than 10% have scaled agents to deliver tangible value, and over 40% of agentic projects are forecast to be cancelled by end of 2027. The failure rate is climbing precisely because the architecture used to build task agents is being stretched to cover decision agents.
The gap is not data quality or model capability. It is causal reasoning. Current LLMs operate on pattern-matched confident guesses: given a history of inputs and outputs, what output is most likely next? That is a question about the past. Decision agents need to answer a different question entirely: what will happen if I take this action, in this specific situation, against these specific constraints? That is a question about cause and effect. No amount of additional historical data converts one into the other. This is the foundational gap. Ergodic fills it.
The Enterprise World Model: foundation layer for decision agents
An Enterprise World Model is a live, executable model of how a system actually operates. It is the thing that lets an agent simulate before it acts, compare alternatives before committing, and trace every recommendation back to a specific, auditable cause. It is composed of two parts.
State is the system's memory: a structured, continuously updated record of every entity, actor, and variable that is decision-relevant right now. Not a historical snapshot, the actual configuration of the operation today. This plant. These suppliers. This inventory position.
Transitions are the relationships that carry the system from one state to the next. For every process there is a function that maps inputs to outputs, not a historical correlation, but a causal description of how the system works. Transitions hold under conditions that have never occurred before because they describe the mechanism, not the history.
Combine continuously updated state with causal transitions and you get the foundational capability: the ability to run the business forward before committing to any action. Set the starting conditions, apply an action, propagate the consequences through the entity graph. The guiding question is simple: given the state of my world today, and the actions I am considering, how will my world wake up tomorrow?
What this unlocks for decision agents
The world model is not a forecasting tool bolted onto an agent. It is the reasoning substrate that makes decision agents safe to deploy at enterprise scale, giving them three capabilities no LLM alone can provide.
This is what separates decision agents that can be trusted at enterprise scale from agents that can only be deployed with a human reviewing every output. The world model doesn't remove human judgment, it makes the agent's recommendations safe enough to be worth examining.
How it works in practice
Every decision-agent interaction follows the same sequence, regardless of domain.
Ground the decision in what is actually true right now. The entity graph and current state place the question in the live configuration of the operation. Nothing downstream is reliable without this step.
Run the operation forward from that state using the transition functions. The agent gets a causal chain, the sequence of downstream effects, not a single extrapolated figure.
Hold the starting state fixed and vary the action. The agent evaluates genuinely different courses of action, each with its own consequences and confidence range.
Trace the chosen action all the way through the network before committing: supplier exposure, margin impact, revenue at risk, inventory obsolescence. This is the shift from reactive exception management to forward planning, surfacing the consequence while the decision is still reversible.
Return a plan that is constraint-consistent and fully auditable, with the causal trace attached. The recommendation is an input to judgment, not a mandate. Decision-makers can inspect it, change the constraints, and ask again.
The world model is not static between cycles. Every prediction is checked against what actually happened, and the discrepancy becomes the signal that sharpens the model. The version that ran yesterday's decision is not the version that runs tomorrow's, each cycle tightens the fit between the model and the operation it describes.
A note on terminology
World models are having a moment. The term was coined by Jürgen Schmidhuber in 1990 and has since become currency across robotics, gaming, video generation, and autonomous driving. The shared commitments are real: simulate rather than extrapolate, and close learning loops against reality. Ergodic shares those commitments and diverges sharply on domain.
Most visible world-model work is spatial, learning the statistical structure of space and time to generate explorable environments or navigate physical scenes. An autonomous-driving world model must obey the laws of physics: a pedestrian steps off the curb, and the model works through vehicle speed, braking, and road grip to determine whether it can stop in time.
An Enterprise World Model must obey the laws of the business. A manufacturer can buy components from one supplier; if that lead time slips, everything downstream is exposed. What it reasons over is the entity graph: who depends on what, and in which direction. The output is a recommended decision, grounded in the operational structure of a specific company, auditable at every step. The architecture transfers because the representational commitments are the same, state, transitions, and a learning loop that closes against reality. The domain is entirely different.
Where this works
The framework is domain-agnostic by architecture. The world model commits to no particular industry, it commits to a representational structure. What makes a domain modelable is constant: the state space must capture decision-relevant variables at the granularity the decision requires; the transitions must be expressible as causal functions that hold under intervention; and the entity graph must encode the constraints the system genuinely obeys, physical, logistical, and contractual laws that data alone cannot enforce.
Where those three conditions hold, decision agents backed by a world model can operate safely. Manufacturing, procurement, pricing, logistics, financial operations: the state space and transition functions change, the reasoning does not.
The path to deployment does not require replacing existing systems wholesale. It begins by identifying the decisions that carry the most consequence, building the world model for those decisions first, validating against historical episodes before anything goes live, and embedding the result as the reasoning engine for the decision agents that operate there. Start where causal reasoning has the highest impact. Build outward.