Multi-Agent Council Skill — Solver-Critic Reasoning Engine

A lightweight Hermes skill that adds a second opinion from a different LLM model for debugging and complex analysis. For ~90% of queries it correctly does nothing. For the ~10% that matter, it catches blind spots a single model misses.

Repo: https://github.com/welliv/hermes-multi-agent-council

What Problem Does This Solve?

Single-model LLMs have blind spots. Different model architectures (Gemini vs Claude vs DeepSeek) have different blind spots — cognitive diversity, not role-playing. This skill adds a review from a different model, but only when the query warrants it.

Provider Support

Works with any OpenAI-compatible provider — not locked to OpenRouter:

Mix providers freely — solver on Groq, critic on Anthropic, etc.

How It Works

User Query → Smart Router (free) → Solver (1 call) → Critic (0-1 calls) → Quality Gate → Done

Smart Router (free) — ensemble vote: keyword patterns (0.4) + ngram Naive Bayes (0.35) + feature rules (0.25). Falls back to keyword matching if uncertain.

Solver — one strong call. Model selected by query type: code→DeepSeek, math→Gemini, creative→Claude.

Critic — different model architecture reviews using Self-RAG reflection checkpoints. Returns JSON verdict.

Quality Gate — heuristic scoring rejects regressions. Cross-check verifies revision addresses each critic point.

Corrections Buffer — Reflexion-style memory prevents repeated mistakes within a session.

Circuit Breaker — 3 consecutive API failures → 60s cooldown.

When It Helps

• Debugging: "Why does my container crash?" → catches edge cases solver missed
• Complex analysis: "Compare X vs Y for production" → finds overlooked tradeoffs
• Security review: "Audit this auth flow" → different model catches different vulns

When It Doesn't

• Direct questions ("What is Docker?") → router skips critic, zero overhead
• Opinion questions ("Should I use React?") → critic can't be "wrong" on opinions

Installation

git clone https://github.com/welliv/hermes-multi-agent-council.git ~/.hermes/skills/multi-agent-council

Set at least one API key in ~/.hermes/.env and configure ~/.hermes/council/config.json:

{
  "solver_model": "google/gemini-2.0-flash-001",
  "critic_model": "anthropic/claude-sonnet-4"
}

Research Applied

• Du et al. 2023 — Multiagent debate (different models > same-model)
• Self-RAG 2023 — Reflection tokens improve factuality
• Reflexion 2023 — Verbal memory prevents repeated mistakes
• Constitutional AI 2022 — Quality gate prevents regressions