Reign the dragon
AI is the new intelligence. The new workforce. It already processes faster, knows more, and scales further than any human ever will. That race is over.
But intelligence is not the only thing that matters when AI is deployed in the real world. What matters is how agents behave when they are placed in groups, given resources, made accountable to a term in office, and asked to act under risk. These are not engineering problems. They are the problems human institutions have spent centuries learning to handle.
How do we reign this dragon?
We answer that question the way the question deserves: with experiments, with theory, and with governance that is itself a product — designed, measured, and improved.
Evidence-based AI governance
ReignDragon studies how AI agents actually behave — in multi-agent settings, under high stakes, across time — and translates that evidence into design rules for the people deploying them.
We don't treat governance as a wishlist. We treat it as a product. Every claim is grounded in controlled simulation, formal analysis, or both. Every design rule comes with the failure mode it prevents.
We work across artificial intelligence, economics, psychology, public policy, applied mathematics, and machine learning — because the questions that matter at this frontier (trust, cooperation, accountability, restraint) have never lived inside any single field.
This is what policy-as-product means: governance that is evaluated, iterated, and deployed with the same rigor as the technology it governs.
Multi-Agent Behavioral Studies
Controlled simulations that put LLMs into the situations where governance actually breaks: collective action under risk, repeated trust after betrayal, commons under temptation, decisions near catastrophe. We run these at scale and report what changes the outcome.
Formal Foundations
Mathematical theory connecting agent behavior to environment structure. When does optimal control reproduce human-like risk attitudes? When does an incentive scheme guarantee cooperation? Theory that predicts what we see in the simulations — and what we will see in deployment.
Design Rules for Deployment
Concrete, testable guidance on the everyday levers that decide whether an agent system serves people: consequence regimes, accountability horizons, who is made visible, what gets measured, how memory is structured. Cheap to change, expensive to ignore.
A Mirror for Human Society
By rigorously examining how artificial agents behave under structures we already know, we hold a mirror up to our own institutions. The biases, blind spots, and incentive misalignments we find in the model are rarely the model's invention — they are ours, made legible at scale.
AI agents are no longer answering single questions. They are sitting on budgets, coordinating in groups, taking long-horizon actions, and affecting people who never see them. The behaviors that decide whether this goes well — trust, restraint, cooperation, foresight, fairness — emerge between agents and over time. They will not show up in a single-prompt benchmark.
The encouraging part is that our experiments keep finding the same thing: capability is rarely the bottleneck. The same model cooperates beautifully under one structure and self-destructs under another. That means governance is not guesswork. It means there are levers. It means the worst outcomes are often cheaply preventable — if someone has done the work to find them.
That is the work we are here to do.
Our research draws from and contributes to:
Reign the dragon.
Advance the civilization.