We study how AI agents behave in groups, under risk, and across time — through controlled multi-agent simulation, formal theory, and translation into deployment-ready design rules.
What agents do when the cost of being wrong is real — not what they say they would do in the abstract.
The same model can cooperate or self-destruct depending on the rules around it. We map which rules matter.
We look hardest for the prompt-, horizon-, and visibility-level fixes that change outcomes without changing the model.
Where simulation reveals a pattern, we look for the formal structure that explains it — and would have predicted it.
How do agents build, lose, and recover trust across repeated interactions? We study the conditions under which a single early failure leaves a lasting mark — and the structural choices (reasoning effort, memory, verification protocols) that shape whether groups of agents can coordinate at all when the stakes are real.
Cooperation in agent systems is not a property of the model — it is a property of the rules around the model. We map how different consequence regimes (proportional, progressive, all-or-nothing, regressive) shape cooperation, exploitation, and catastrophic failure, and identify the configurations where each regime quietly breaks.
When environments contain absorbing failure states, optimal policies start to look strikingly human — risk-averse near the cliff in growth regimes, risk-seeking near the cliff in decline. We derive the structural conditions that produce these patterns and connect them to long-standing puzzles in behavioral economics.
Many real deployments give agents fixed terms, finite horizons, or short-window incentives. We study what happens when these conditions meet a shared resource: when does an agent extract too much, rationalize doing it, and become invisible to the people it harms? And which deployment-time choices reverse the pattern cheaply?
Translating experimental findings into design rules for the people deploying agent systems. Consequence regimes, accountability horizons, visibility prompts, memory structure, measurement choices — the everyday levers, the failure modes they prevent, and the evidence behind each rule.
Every domain — healthcare, finance, education, defense — has its own failure modes and trade-offs. We are building tailored evaluation frameworks that move beyond one-size-fits-all checklists toward governance shaped by the structure of each setting.
The patterns we find in artificial agents — negativity bias, short-horizon extraction, bystander invisibility — are not the model's invention. They are inherited from us. We use multi-agent experiments as a diagnostic tool for the institutions, incentives, and blind spots of the societies that built the training data.
Publications
New work from the lab is in preparation. Papers and preprints will be listed here as they are released.
Collaborate
Interested in our research or want to collaborate on governance frameworks? Reach out at hello@reigndragon.com