You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$19,900 left to reach the funding goal.
I’m seeking $5,000–$20,000 to study how agent personality traits may shape the societies they build, and how those societies in turn shape their personalities. I’ll apply naturalistic, random-ish personality steering, establish initial tribes of diverse agents in the game One Hour One Life (OHOL), and monitor them across generations. I’ll test whether an agent’s steered traits affect their survival, role, and descendants, and whether social and environmental pressures measurably push those traits to persist, drift, or converge.
I think that diverse agent societies are:
more favorable for alignment than tiling the world with one agent (echoing Landemore’s case for cognitive diversity),
somewhat inevitable,
and understudied.
OHOL works well as a small, controllable version of the agent societies we’re starting to build. There’s an implicit need to survive, strong benefits for cooperation, and opportunities to specialize; and while the long-term goal of civilizational flourishing is broadly clear, the game’s depleting resources require societies to keep adapting.
This naturalistic and large setting allows us to ask a lot of questions. I’d like to start with open-ended ones and use what I observe to design more focused tests.
How do steered traits affect who survives, who leaves descendants, and who takes on roles like exploring, caregiving, or coordinating others?
Do findings with agents mirror findings from humans playing the same game?
How do scarcity and interaction with other agents affect whether steered traits persist, drift, or converge? Do agents start behaving more like the agents they spend the most time with?
Setup
With a few exceptions, most of the details of the game will be kept true to the original.
I’ll extract persona vectors for several traits tied to the game’s tradeoffs, check that they change behavior outside the game, and confirm steered agents can still play competently. Children will be regularly spawned to mothers and inherit their mother’s steering coefficients with small mutations. The main study will run large populations across independent worlds, comparing different trait mixes with unsteered groups.
Related work
Closely related agent-society studies give agents personalities through prompts (Liao et al., 2026; Why Are We Moral?), but prompted personas often fail to take hold or fade over long interactions. Steering allows finer control over personality and lets me measure drift in its activations. It may also help explain how traits spread: models seem to represent others with some of the same internal machinery they use for themselves, both on the assistant axis and in Anthropic’s emotion vectors.
This project strongly complements the amazing Mindscapes! It differs in the setting and the survival and selection mechanisms, but I’m 100% keeping an eye on their Persona Geology findings, and I agree with their evaluation of “qualitative, non-prescriptive character.”
There are a few cool extensions to pursue here.
What happens to different personas under scarcity? I’ll also track emotion vectors for states like desperation and calm, to see whether some traits are more prone to desperation and whether it predicts hoarding or abandoning kin.
How do personas influence larger and more developed economies? Anthropic recently released Project EVE, studying lone Claude traders and small teams of agents. Individual traders mostly arrived at the same strategy, and teams reliably sorted themselves into roles. I’d love to scale this experiment up and let different societies of agents interact in such a complex game.
How do personas shape strategy in adversarial/scheming settings? In The Castle Doctrine, by Jason Rohrer, the creator of One Hour One Life, players build traps to defend their homes and break into each other’s houses.
I’ll call this a successful pilot if steered agents reach adulthood at rates close to unsteered agents and their traits differ measurably on behavioral probes. Beyond that, I’ll write up a null result (and publish code, logs, etc.), though once I’ve ensured traits cause meaningful behavioral changes, I expect most other failure modes to still be interesting.
This funding would cover my own stipend so I can keep working on the project full time until the Anthropic Fellows Program starts in January, if I’m admitted (I’m in the interview stages), or longer if I continue independently. Part of the draw is that this kind of experiment scales well, so everything besides a stipend goes into that: more or different trait adjustments, more independent worlds, more generations, etc., letting me pursue unexpected findings.
Just me so far! I’ve published some hobby projects on personality steering, and at Theorem I post-trained Kimi K2.6 agents to be better at formally verifying properties of software. I’m getting very casual guidance from researchers at frontier labs, but I’d also very much welcome offers of more regular check-ins from senior researchers in this area.
I’ve received $10,000 for this project from Andrew Gritsevskiy’s Promontory Labs. I’m also applying to BlueDot Impact and the Transformative AI Fund for career development support more broadly, since this is my first major foray into personality work :)