Predictive intelligence through agent-based population simulation.
Website · Announcement · CLI Reference · Architecture
Extropy creates synthetic populations grounded in real-world distributions, connects them in social networks, and simulates how they respond to events — each agent reasoning individually via LLM.
Simulate anything: Policy changes. Pricing decisions. Product launches. Crisis response. Any scenario where humans form opinions, make decisions, and influence each other.
Caution
"Extropy is still in active development and behavior may change between versions."
For reliable execution and triage, we strongly recommend running Extropy through an agentic harness (for example: Codex or Claude Code) rather than manual ad-hoc CLI usage.
Simulation can be expensive at scale (especially high-fidelity, multi-timestep runs). Start with small runs first; we recommend beginning at around 500 agents before scaling up.
pip install extropy-run
export OPENAI_API_KEY=sk-...
# or: ANTHROPIC_API_KEY=... / AZURE_API_KEY=... + AZURE_ENDPOINT=...Requires Python 3.11+. uv recommended.
# Create study folder and build population spec
extropy spec "Austin TX commuters" -o congestion-tax -y
cd congestion-tax
# Create scenario with events and outcomes
extropy scenario "Response to $15/day congestion tax" -o congestion-tax -y
extropy persona -s congestion-tax -y
# Sample agents and generate network
extropy sample -s congestion-tax -n 500 --seed 42
extropy network -s congestion-tax --seed 42
# Run simulation
extropy simulate -s congestion-tax --seed 42
# View results
extropy results
extropy results segment income- Population — LLM discovers attributes, researches real-world distributions, samples agents
- Network — Builds structural + similarity edges; edge types affect information flow
- Two-pass reasoning — Agent role-plays reaction, then classifier extracts outcomes
- Propagation — Opinions spread through network; agents update after hearing from peers
| Feature | Description |
|---|---|
| Population | |
| Any geography | US, Japan, India, Brazil — define attributes with your distributions |
| Real grounding | LLM researches actual demographics, cites sources |
| Household mode | Correlated partners, NPC dependents, assortative mating |
| Agent focus | Primary adult, couples, or full families as reasoning agents |
| Network | |
| Structural edges | Partner, household, coworker, neighbor, congregation, school parent |
| Similarity edges | Acquaintances and online contacts from attribute similarity |
| Small-world | Calibrated clustering coefficient and path lengths |
| Simulation | |
| Two-pass reasoning | Role-play first, classify second — reduces schema anchoring in outcome extraction |
| Conversations | Agents talk to each other; both update state independently |
| Memory | Persistent memory traces with fidelity-based prompt trimming |
| Conviction | Affects sharing probability and flip resistance |
| THINK vs SAY | Internal monologue separate from public statement |
| Timeline events | New information injected at specified timesteps |
| Outcomes | |
| Categorical | Known decision space (buy/wait/skip) |
| Boolean | Binary decisions (will share, will switch) |
| Float | Intensity measures (sentiment, likelihood) |
| Open-ended | Free text — discover categories post-hoc |
git clone https://github.com/exaforge/extropy.git && cd extropy
pip install -e ".[dev]"
pytest