Ask an AI assistant a question and it will, nine times out of ten, agree with you. That's lovely for the ego and useless for decisions. The interesting question was: can you get genuinely better answers by making several AI models disagree with each other on purpose?
This one was our own itch. We'd been building agents and automations for clients and wanted a public, clickable answer to "what do you actually mean by AI agents?" — something better than a slide deck. So we built The Council, based on Andrej Karpathy's LLM Council idea.
Six advisors, each given a deliberately narrow personality: The Contrarian hunts for the fatal flaw, the First Principles Thinker questions the question, the Expansionist looks for the upside, the Outsider answers with zero context, the Executor only cares what you do on Monday, and the Trend Scout goes off and researches what people are actually saying online right now.
Each advisor answers independently — they can't see each other's responses, which is the whole point. Then their answers are peer-reviewed anonymously, and a Chairman model synthesises everything into a verdict that has to separate "things multiple advisors agreed on independently" from "genuine disagreements worth your attention".
Under the hood it's a Next.js route streaming NDJSON over the Vercel AI Gateway — cheaper, faster models for the advisors, a stronger one for the Chairman, because the synthesis is where the quality lives. Honest trade-off: a full Council run costs more and takes longer than just asking a chatbot. That's the price of an answer that's been argued about.
It's live, it's free, and you can try it now. People use it for pricing decisions, career moves, and at least one wedding speech dilemma. It's also the most effective sales tool we've got for AI consultancy work — because when someone asks whether we can build multi-agent systems, we just send them a link.
