Skip to main content

Agents & Personas

An agent is one participant in the simulation. Every agent has an identity, a set of environments it can act in, an evolving belief state, and a short memory of its recent actions. During a run, an agent observes its environments each round and uses an LLM to decide what to do — post, reply, follow, vote, trade, or do nothing.

From seed to cast​

Agents are generated from the seed document. An LLM call asks the smart model to list the entities most relevant to the prediction goal and returns them as structured records — each with a name, a type (for example a person or an organization), and a short summary. Every extracted entity becomes one agent: the engine seeds the agent's persona from the entity's name and summary so the agent acts as that participant from its very first round.

If the LLM extraction can't be parsed, the pipeline falls back to a simpler capitalized-word scan of the seed so a run always has at least one agent.

What an agent carries​

  • Persona — a short description of who the agent is, built from its source entity.
  • Environments — the social and/or market environments it participates in.
  • Belief state — its positions, confidence, and trust in other agents, which changes every round (see Beliefs & Stance).
  • Memory — a rolling window of its most recent actions, included in its context so behavior stays coherent across rounds without growing unbounded.

Personas in the results​

The seed-time persona is what drives behavior during the run. Separately, after the simulation finishes, an LLM-backed step can rewrite each agent's profile into a richer two-to-three-sentence persona based on what the agent actually did — its sample posts and sentiment arc — replacing the default one-line activity summary shown in the results. This is a descriptive, post-hoc enrichment; if it fails for an agent, the one-line summary is used instead.