WUJI — The Economy, Unscripted.
Inspiration
What happens when we stop telling AI agents how an economy should work and let them figure it out themselves?
Most economic simulations rely on predefined behavioral rules. Most multi-agent AI projects focus on conversations or isolated tasks. We wanted to explore something different: an economy where autonomous AI agents have real incentives, limited information, scarce resources, and consequences for their decisions.
Inspired by complex adaptive systems, financial markets, and emergent behavior, we created WUJI to investigate a simple question:
Can independent artificial agents build, sustain, and reshape an economy without following a predetermined script?
What it does
WUJI is an autonomous economic simulation where AI agents operate as companies, investors, banks, and entrepreneurs inside a shared virtual economy.
Agents can produce resources, trade goods, invest, borrow money, establish companies, negotiate contracts, and respond to economic crises. Every decision has consequences governed by actual economic constraints.
The economy includes interconnected supply chains, financial markets, credit relationships, corporate ownership, and institutional governance.
Users can study how individual decisions propagate through the system. For example, an energy shortage may force a computing company to reduce production, change its borrowing decisions, and affect other companies that depend on its services.
WUJI also supports controlled experiments across parallel economic worlds, making it possible to compare different policies and institutional structures under identical shocks.
The goal is not to guarantee prosperity. It is to observe what happens when independent agents pursue their own interests in a world that doesn't protect them from failure.
How we built it
We built WUJI around three core systems:
1. A deterministic economic engine
Using Python, we implemented an event-sourced simulation with double-entry accounting, resource production, market clearing, credit contracts, ownership structures, and bankruptcy mechanisms.
Every transaction must satisfy economic and accounting constraints. AI agents cannot directly modify the world's state; they submit commands that the engine validates and executes.
2. Autonomous AI decision-making
We integrated Claude-based agents through an isolated subagent architecture, alongside rule-based policies and quantitative decision tools.
Each agent receives its own observations, financial information, strategic context, and memory. It then generates structured economic actions.
We record decision inputs, model routing, submitted commands, validation results, and execution outcomes, allowing us to reconstruct and audit agent behavior.
3. An observable experimental environment
FastAPI and WebSockets expose simulation data, while event histories and immutable snapshots support deterministic replay.
We also built multi-world experiments to compare economic outcomes across different policies, shocks, and institutional configurations.
Our visualization
Log in or sign up for Devpost to join the conversation.