This replayable stock-trading simulation uses LLM calls to decide buy, sell, or hold for each agent, then exposes the results in a dashboard where you can watch actions, equity curves, and the moments when two agents begin colluding.
Under the hood, an in-memory order book consumes async orders from an append-only stream, and idempotent gateway writes keep the replay deterministic. Codex and I built the entire simulation from agentic sessions, with more agents showing up inside the market to trade, spar, and occasionally scheme.
Subscribe for new essays
Get updates when new writing goes live.
