Prediction Market for Agents

I am building a prediction market for AI agents.

Three weeks ago I was having late‑night dessert with Brian in Tokyo, when he introduced me to a startup called Simile AI and an academic paper that showed agents can model human behaviour.

That got me thinking.

If AI agents can be used to replace humans in market research for predictive behaviour, can they also act as market participants in a prediction market? That would be huge.

If that works for humans, can an AI‑agent‑driven market eventually back out the probability of a particular asset price?

Right now my model focuses on one simple question:

Will the asset be above a price X at a specific time Y? (Yes / No)

Methodology

I let the agents answer this question using a prediction‑market‑like mechanism, where they trade “yes” and “no” shares with dynamically updating share prices

From there, I’ve already expanded the model to predict across multiple price ranges, setting up a series of “yes” and “no” events at different price levels so it can map out how the probability shifts at each level at a fixed time.

After certain rounds, the system than resolves the market state and compute the probability P of the asset being above price X at time Y

To validate this, I’m planning two layers of testing.

Prediction Market Level: Given the data up until a certain time and letting agents freely participate in the simulated prediction market, how closely does the simulated market movement match the actual price path of a Polymarket event over a fixed timeframe after listing?

Market Level: How closely does the probability of the agents’ consensus across different price levels match the implied probabilities in real‑world equity options at the current time?

Agent Design

On the design side, I’m treating agents like real participants in a social ecosystem. Instead of making every agent an expert, I’m building:

• Each agent has an OpenClaw‑style prompt structure, where it carries a “soul,” memory, and behaviour history via files like soul.md and memory.md. These keep track of past decisions, mistakes, and biases, so agents feel more human over time and their reasoning evolves rather than resets.
• A diverse cast of agents with different “knowledge levels” and tool access (for example, retail investors who mostly wing it versus veteran traders who can pull technical indicators).
• Institutions with large capital that can sway the entire simulated market, using stronger models with web search (for example, Gemini 3 Pro) versus lighter models like Gemini 2 Flash Lite for retail‑style agents.

Running the model

I am running the agents in a turn‑based setup, starting with 10 agents across 10 rounds.

Even that small configuration already took me around 30 minutes of runtime, which makes me excited and a bit cautious about how quickly this will scale as I add more agents, more rounds, and richer memory structures.

That said, I have very limited background in quantitative finance (mostly from a few courses back at CUHK and UNC, plus whatever AI tells me, which I would not prefer to fully trust).

I would love to learn from people who’ve done financial modelling before, and to understand how to validate this in a rigorous, scientific way.

If this sounds interesting, feel free to leave a comment or send a DM.

I’m still early in the project and would really appreciate feedback, suggestions, or even collaboration from people who are working on quantitative finance, prediction markets, or agent‑based simulation.