AgentWorld · Human behaviour research

A living model of human behaviour.

We simulate thousands of people, each with a body, habits and traits measured from real behaviour, to see how societies respond to change before it happens.

Scroll to meet one person
01 · One person

It starts with one person.

Not a persona written in a prompt. A set of traits measured from what they actually do: how they weigh losses, discount the future and learn from mistakes.

02 · Individuality

No two people are alike.

Each simulated person is fitted individually, so a population keeps the real spread of behaviour, including the tails where crises begin.

03 · The body

People have bodies.

Stress, fatigue and scarcity rise and fade over days, and they change every decision. Our agents carry that state with them.

04 · Habit

Mostly habit. Sometimes thought.

People run on routine and stop to think only when something surprises them. Our agents do the same, shown in blue. It makes them more realistic, and it lets us simulate millions.

05 · Society

Then thousands of them, together.

Neighbourhoods share one world with markets, news and each other. Every action changes what everyone else sees.

06 · A shock

Change one thing. Watch who tips first.

Raise prices and the households with the thinnest buffers move first, stopping to rethink. Their stress travels through the market to everyone else. Scroll slowly to scrub the cascade.

07 · Reality

Then we check it against reality.

Every prediction is scored on behaviour the model never saw, against how consistent people are with themselves. When we miss, we publish that too.

The difference

Today's simulated people only talk. Ours behave.

Language-model agents predict what people say in surveys. Real behaviour depends on the body, on habit, on memory and on the people around us. We build all four back in.

Persona agentsAgentWorld agents
Learned fromWhat people sayWhat people do
IndividualsDescribed in a promptFitted from behaviour, one by one
BodyNoneStress, fatigue and scarcity over time
DecisionsEvery step deliberatedHabit first, deliberation on surprise
MemoryStores everythingConsolidates, forgets and drifts
ProofSounds plausibleScored against held-out behaviour
What we solve

Questions no survey can answer.

Each one is about how a population behaves when its world changes. Choose a question to run a simplified version below.

Playground

4,000 people. Your shock.

Baseline. Everyone is following their habits.
People0
On habit0%
Rethinking0%
Mean stress0.00
Stress, last 20 s

A simplified model running in your browser. Each person has their own financial buffer, stress level and routine. These are illustrations of the mechanism, not forecasts.

Method

Plausible isn't enough. We measure.

Behavioural accuracy
model − chance human − chance

People aren't perfectly consistent with themselves. So we score a model against how well each person predicts their own behaviour two weeks later. A score of 1.0 means the simulation knows a person as well as they know themselves.

  • Behaviour, not wordsChoices, transactions and actions. Survey answers are never the target.
  • Data the model never sawTested on new people and new periods.
  • Committed in advanceForecasts are fixed and published before outcomes exist.
  • Misses includedEvery result ships with every baseline, including the ones we lose.

Understand people before the world changes them.

We're a small research team building the instrument. We work with cognitive scientists, neuroscientists, economists and institutions that hold consented behavioural data.

Get in touch

Tell us who you are and which question you'd like to study. We reply to every message.

ResearchersDecision science, computational psychiatry, neuroscience and agent-based modelling.
Data partnersConsented, longitudinal behaviour across real shocks.
Public-interest teamsPopulation forecasts with honest error bars.