展示HN：Boardroom MCP - 面向AI代理的多顾问治理引擎
嗨，HN，
我注意到在构建自主代理时存在一个反复出现的问题：它们在执行方面表现出色，但在细致判断上却很差。当面临模糊情况时，它们只是幻想出最具统计概率的路径，而没有考虑二阶效应。
我没有试图通过大量系统提示来解决这个问题，而是构建了一个MCP（模型上下文协议）服务器，将决策转交给一个多顾问的“董事会”。
它的工作原理如下：
1. 代理遇到决策 -> 调用 `analyze()` MCP 工具。
2. 服务器将查询路由到相关顾问（来自38个领域，450多个个人资料）。
3. 顾问们进行辩论。这是核心机制：如果顾问们达成一致的速度太快，系统会标记此情况。紧张感是必需的。
4. 服务器综合辩论结果，将其记录到持久的账本中（以便代理拥有机构记忆），并返回一个风险评分的建议。
我在本地构建了这个系统，因此没有任何云依赖，您的代理数据仍然属于您。它可以与Claude Desktop、Claude Code、Cursor、Windsurf或任何MCP客户端原生工作。
您可以在这里查看核心引擎： [https://github.com/randysalars/boardroom-mcp](https://github.com/randysalars/boardroom-mcp)
如果您想阅读文档： [https://salars.net/boardroom/docs](https://salars.net/boardroom/docs)
我很想知道是否还有其他人尝试过结构化的多代理辩论与LLM作为裁判的模式。
查看原文
Hey HN,<p>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.<p>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."<p>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.<p>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.<p>You can checkout the core engine here: <a href="https://github.com/randysalars/boardroom-mcp" rel="nofollow">https://github.com/randysalars/boardroom-mcp</a>
If you want to read the docs: <a href="https://salars.net/boardroom/docs" rel="nofollow">https://salars.net/boardroom/docs</a><p>Would love to know if anyone else has experimented with structured multi-agent debate vs LLM-as-a-judge patterns.