EP4: The Point of No Return
Why a single conversation can permanently collapse a relationship. The hidden psychology behind proposals, rejection, and the cliff.
Read more
ReignDragon
Reign the dragon
AI is the most powerful force humanity has ever created. ReignDragon is a research lab studying how AI agents behave under pressure — in groups, under risk, across time — and turning that evidence into governance that works.
What we do
01
Behavioral Experiments
We run controlled multi-agent simulations where LLMs face high-stakes decisions — cooperation under risk, trust under uncertainty, commons under temptation — to surface the failure modes that don't appear in single-turn benchmarks.
02
Formal Theory
We connect agent behavior to mathematical structure: when do optimal policies look risk-averse, when do bounded horizons induce extraction, when does an environment guarantee cooperation? Theory that predicts deployment.
03
Policy-as-Product
Findings become design rules. Consequence regimes, accountability horizons, visibility prompts, memory structures — the everyday levers that decide whether a deployed system serves people or quietly harms them.
Why it matters
Agents are leaving the sandbox
LLMs are no longer answering single questions. They are coordinating in groups, holding budgets, taking actions across long horizons, and affecting people who never see them. The behaviors that matter now — trust, restraint, cooperation, foresight — emerge between agents and over time, not in any one prompt.
Structure beats sentiment
Our experiments keep finding the same thing: capability is not the bottleneck. The same model cooperates or self-destructs depending on consequence design, accountability horizon, and who is made visible. These are governance choices, and they are cheap to fix — if we know to fix them.
Our story
ReignDragon began with a simple observation: every serious account of AI risk eventually becomes an account of incentives, institutions, and human nature — topics economics and psychology have studied for a century, but that the AI field keeps re-deriving from scratch.
We built a lab to close that gap. We design experiments that put language models into the situations our institutions were built to handle — collective action, repeated trust, decisions near catastrophe, fixed terms of office — and we measure what actually happens, episode by episode, at scale.
What we find is consistent and useful: AI agents inherit recognizable patterns from the data they were trained on, and they break in recognizable ways when the structure around them is wrong. That is both a warning and a gift. It means governance is not guesswork. It means there are levers, and they can be pulled.
Latest
Signal
Reign the dragon.
Advance the civilization.