I noticed a recurring issue when building autonomous agents: they're great at execution but terrible at nuanced judgment. When faced with ambiguity, they just hallucinate the most statistically probable path without considering second-order effects.

Instead of trying to fix this with massive system prompts, I built an MCP (Model Context Protocol) server that offloads decisions to a multi-advisor "boardroom."

How it works: 1. Agent encounters a decision -> calls the `analyze()` MCP tool. 2. Server routes the query to relevant advisors (from 38 domains, 450+ profiles). 3. Advisors debate. This is the core mechanic: if advisors agree too quickly, the system flags it. Tension is mandatory. 4. The server synthesizes the debate, logs it to a persistent LEDGER (so the agent has institutional memory), and returns a risk-scored recommendation.

I built this locally so there's zero cloud dependency and your agent data stays yours. It works natively with Claude Desktop, Claude Code, Cursor, Windsurf, or any MCP client.

You can checkout the core engine here: https://github.com/randysalars/boardroom-mcp If you want to read the docs: https://salars.net/boardroom/docs

Would love to know if anyone else has experimented with structured multi-agent debate vs LLM-as-a-judge patterns.