Every major decision I make at AC Solutions goes through a council of five AI agents before I act on it. This is not a gimmick. It has replaced the three strategy consultants I used to pay $800/hour combined, and it is more useful.
Here is how I got there and how it works.
In early 2024 I was using ChatGPT the way most executives do: as a smart search engine. Ask a question, get an answer, feel productive. The problem is that a single AI model gives you a single perspective. It will not tell you you are wrong. It will not challenge your assumptions. It is fundamentally a yes-machine unless you deliberately push back.
I read a paper on multi-agent debate in LLMs and had a simple thought: what if the agents argued with each other? Not took turns summarizing, but genuinely debated from assigned expert roles?
I built the first version in a weekend using the OpenAI API. Five agents: CFO (financial risk focus), CMO (growth and marketing), CTO (technical feasibility), Investor (exit and valuation lens), Devil's Advocate (whose only job is to destroy the idea). Each agent gets the same question, drafts a position, then responds to the others. After three rounds, Gemini 2.5 Pro synthesizes the debate into a structured decision document.
The first real test: should we drop one of our low-margin CPaaS routes to cut costs? Single-model answer: yes, margin is thin. Multi-agent debate: the Devil's Advocate pointed out that route was gateway-pricing for three enterprise clients who would churn if we dropped it, and the Investor agent calculated that the retention value was 14x the margin loss. We kept the route. We kept the clients.
The model routing matters as much as the architecture. I use DeepSeek-chat for the expert agents — cheap, fast, smart enough to argue effectively. Gemini 2.5 Pro with thinking mode enabled for synthesis. o3 for financial calculations when exact numbers matter. The whole council costs less than $1 per session.
I have now run over 50 council sessions. Pricing decisions, hiring choices, product roadmap prioritization, whether to raise or stay bootstrapped. The pattern I noticed: the council is most valuable when I am confident I already know the answer. That confidence is usually the danger zone.
The framework is open source on GitHub under debate-engine. You do not need my stack to use it. You need five system prompts, an orchestrator, and a willingness to hear things you do not want to hear. That last part is the hardest.