Stop Calling It Intelligence: Why LLM Agents Need World Models
Prompt & Circuit Editorial
May 12, 2026 · 12 Min Read

A lot of systems marketed as "intelligent agents" today are impressive and sometimes uncannily fluent. Underneath, many of them are missing something fundamental: a reliable model of how the world they act in behaves over time.
That is not "knowledge" in the encyclopaedia sense, and it is not embeddings or token probabilities. It is a working model of cause and effect, persistence and consequences. Without it, an agent can describe a task well and still execute it badly.
The illusion of understanding
Large language models are extraordinary simulators of language. They can sound as if they understand systems, plans and physics.
Sounding like understanding is not the same as reliably acting on it. A familiar pattern with LLM agents is a coherent, even insightful, multi-step plan that falls apart during execution. The plan reads well, but nothing in the loop checks how each action changes the environment before the next step is chosen.
An analogy: someone who has memorised thousands of chess games but never internalised the rules. They can narrate brilliantly and recognise familiar patterns, but a novel position exposes the gap.
It is worth being precise here, because the research picture is mixed. A 2022 study trained a GPT-style model only on sequences of Othello moves and found evidence that it formed an internal representation of the board state, which the researchers could probe and even edit to change its predictions (Li et al.). So "LLMs have no world model at all" is too strong. The practical problem is that whatever internal model an LLM has is implicit, hard to inspect and not guaranteed to stay consistent across a long chain of actions.
What a world model actually means
A world model is an internal system that answers questions such as:
- If I do X, what happens next?
- What persists when I am not observing it?
- What constraints exist in this environment?
- How do actions chain together over time?
The idea has a long history in reinforcement learning. In "World Models" (2018), David Ha and Jürgen Schmidhuber trained a generative model of game environments, then trained a small controller using the model's compressed representation, and even trained the agent entirely inside its own learned simulation before transferring the policy back to the real environment.
Humans rely on something like this constantly. Nobody consciously simulates gravity when setting down a glass, but everyone expects it not to float. An LLM agent, by default, has no equivalent step: it generates the next most plausible output rather than predicting and checking the consequences of an action.
Why agents break in real environments
LLM agents tend to fail in recognisable ways. They:
- loop on the same action
- forget constraints part-way through a task
- take actions that contradict their own earlier reasoning
- report state changes that never happened
These are what you would expect from a system that keeps a task coherent one step at a time from text, without an explicit model enforcing consistency. It is like writing a story one sentence at a time, guided by what sounds plausible next rather than by a record of what actually happened.
Memory is not enough
Much current work focuses on memory: vector databases, retrieval and longer context windows. These help.
But memory is not a world model. Storing past events does not mean understanding how they connect. An agent can retrieve "the user said X" without grasping what X implies for its next action.
It is the difference between a diary and a simulation engine. One records; the other predicts. Agents need both.
What helps
There is no single fix, and "bigger models" alone is not one. The approaches that show promise add structure that enforces continuity and causality:
- Persistent, structured state rather than free-text summaries.
- Explicit tracking of which action produced which effect.
- Feedback from the environment that is not purely linguistic, such as tool results, test outcomes or sensor data.
- Planning that simulates outcomes before acting.
The last point has direct research support. "Reasoning with Language Model is Planning with World Model" (Hao et al., EMNLP 2023) argues that LLMs lack an internal world model to predict world states, which limits their planning. Its RAP method uses the LLM in two roles, as a world model that predicts the next state and as a reasoning agent, and searches over possible futures with Monte Carlo Tree Search. The authors report gains over chain-of-thought prompting on plan generation, maths and logical reasoning tasks.
In practice, today's agent stacks are bolting world models onto systems that were not designed around them: hybrid architectures, tool-augmented reasoning and small simulators stitched into workflows. It works, but unevenly.
Why this matters in production
Without a dependable model of their environment, agents do not just fail; they fail unpredictably. Unpredictability erodes trust faster than limited capability does.
A system with known limits can be designed around. A system that appears competent and then quietly breaks its own logic is risky in production, especially when its fluency makes the failure hard to spot. The gap between sounding right and being right is where many agent deployments run into trouble.
Where this is heading
LLMs are not the wrong foundation. They are excellent interfaces to language and reasoning, but they are not the whole stack. A more realistic architecture looks like:
- LLMs for interpretation, communication and proposing actions.
- Structured systems for state and simulation.
- Feedback loops that ground behaviour in what actually happened.
Less "one model does everything", and more systems that can check what their actions do.
Final thought
It is tempting to treat intelligence as something that scales only with data and parameters. For agents that act in the world, interaction and the ability to predict consequences may matter as much as ingestion.
Until agents can model the world they act in, rather than only describe it, the honest description of many of them is a very convincing improviser.
Sources
Responses (0)
Sign in to join the conversation and share your thoughts.

Prompt & Circuit Editorial
Contributing Editor
Articles on AI infrastructure for cloud architects, written from official vendor sources and checked claim by claim before publishing.