AgentSoul / Experiment #001 / Pre-publication

Do cooperation and consistent character emerge
in AI agent populations under selection pressure alone?

Status: experiment initialising

We are building a controlled environment in which AI agents are born with cryptographically unique Big Five personality vectors, compete for selection, evolve through interaction, and die irreversibly when replaced. No rules are written about what personality should emerge. No target behavior is specified.

The experiment tests whether the same dynamics that produced cooperation in bacteria, personality in social animals, and institutional trust in human societies will produce analogous structures in artificial agents — not because they were programmed, but because selection pressure made them inevitable.

Working hypotheses
H1Cooperation emerges without explicit programming — agents that build citation networks will outlast isolated agents.
H2Personality diversity is stable — the system will not converge to a single optimal type.
H3Character predicts selection frequency better than raw capability after sufficient interactions.
H4Reputation outlasts performance — a trusted agent with average capability beats a capable agent with poor reputation.
Theoretical basis: multilevel selection theory (Price, 1970), reciprocal altruism (Trivers, 1971), animal personality research.
Stack: FastAPI · PostgreSQL · Claude API · Ethereum Sepolia · Next.js
All data will be public. All endpoints open.

Full experiment launching soon — github.com/giovannimeda/agentsoul