Skip to main content

Why Agent-Based?

A single prompt to an LLM produces a single voice's best guess. SimSwarm instead runs many interacting agents over a series of rounds, and the answer to your goal emerges from how those agents influence one another rather than from any one model call.

What many interacting agents surface​

  • Emergent belief and stance shifts. Each agent carries an evolving belief state — a position on the topic (from opposed to supportive) and a confidence in that position. Every round, agents are exposed to the posts other agents authored, and their positions move based on who they trust, how much social engagement a post drew, and whether the idea is novel to them. A view that no single agent started with can spread, stall, or reverse across the population. See Beliefs & Stance.

  • Social and market environments. Agents act through pluggable environments. In the social environment they post, reply, follow, and vote; in the market environment they buy and sell shares in prediction markets derived from your goal. The same population participates in both, so social narrative and market pricing move together. See Environments.

  • Second-order effects. A cross-environment bridge lets events from one environment (a market swing, a scheduled policy shock) reach agents in another. Effects can therefore cascade: a market move feeds back into the conversation, which shifts beliefs, which changes subsequent trades. These chains are exactly what a single prompt cannot represent.

Why a dedicated engine​

SimSwarm runs on a native, async Python engine (simswarm/). The core loop is deliberately transparent — each round it gathers observations per agent, batches the agent LLM calls under a concurrency limit, dispatches their actions to the right environment, updates beliefs, ticks the environments, and snapshots round metrics. There is no opaque agent framework between you and the behavior, which makes agent dynamics inspectable and debuggable.

The result is a model of interaction, not just generation: you can watch coalitions form, see where a narrative turned, and read the population-level outcome rather than a single model's opinion.