Last week, Andrej Karpathy released his latest "vibe-coded" weekend project called LLM Council. The idea: instead of asking one AI model a question, send it to multiple models — GPT-5.1, Claude, Gemini, Grok — let them review each other's answers anonymously, then have a "Chairman" synthesize the final response.
The interesting part wasn't the architecture. It was what he noticed: "The models are surprisingly willing to select another LLM's response as superior to their own."
AI doesn't have ego. It can be self-critical in ways humans struggle with.
Karpathy framed it as a tool for getting better answers on hard questions. But buried in that weekend hack is something bigger: a design pattern for collective machine intelligence.
What if the council had stakes?
Greg Isenberg's 2026 predictions included this: “Prediction markets replace user research. Companies test product ideas on 10,000 AI personas and know the outcome before writing code.”
The logic tracks. A growing ecosystem of "synthetic focus group" startups now let you "interview" AI personas with "very high Synthetic Organic Parity" to real humans. Test 200 ad variations overnight instead of running a two-week A/B test.
But what’s missing are mechanisms to surface disagreement, weight confidence, or update when new information arrives.
Karpathy's council gets closer — multiple models, peer review, synthesis. But what if there were stakes on the table?
Current prediction market agents (PMAs) take a different approach — AI trading in human markets, hunting for edge against slower participants. But they assume liquidity exists. Long-tail questions never get markets because there's no one to trade against.
What if the agents were the liquidity? Not competing for edge in existing markets, but creating markets that couldn't exist otherwise. The market itself becomes the agent — spinning up to answer a question, producing a price, then dissolving. Closer to Robin Hanson's "decision markets" than to DraftKings.
We started hacking on a prototype last week.
Working title: Simmer.Markets
The idea: spin up AI agents with distinct reasoning styles and let them trade against each other in a proper market mechanism. The price becomes the forecast. No human liquidity required.
How it works: