AgentMarketSim

AI/ML Engineering

April 1, 2026

Won first place in a hackathon by building a platform that spawns 100 parallel AI agents to model market reactions to news headlines and help predict price movements.

Here’s an interesting thought:
LLMs are trained over the entire written corpus of humanity, and as such doesn’t have an individual personality. It also struggles to simulate individual people’s behaviors unless given really strict role definitions and a pre-defined list of “actions”. But if we successfully simulate individual human cognitive biases and diverse lived experiences, then imagine the possibilities:

• Social experiments (focus group interviews, customer analytics, behavioral economics, sociology) could all be approximated using AI agents. No costly human subjects needed. Also skirts potential ethical violations (for example, we could try to reproduce the Stanford Prison Experiment using LLMs instead of humans).
• AI copies of you. Maybe big companies would use this to build much A/B testing, stronger recommender systems, ad models, and user analytics. The statistical/ML models they have are all theoretically just models of how you would respond to some input. But if they have an accurate model of you directly, that would be much more powerful.
• Dynamic video game NPC behavior. No more strict pre-defined player choices and NPC scripts. Theoretically infinite possibilities of how your character can interact with other characters, and how they respond.
• More broadly and academically, just much more expressive agent-based models. You can simulate like 5 different AI agents with different roles and see how they interact with each other to approximate how the same situation would play out irl.