Hermes Atlas
Multi-agent & orchestration

PolyGnosis

mosesman831/PolyGnosis

Adversarial multi-model consensus skill for Hermes Agent: independent solving, critique and formal scoring

In short

PolyGnosis is an adversarial multi-model consensus skill for Hermes Agent that has several models solve a problem independently, critique each other and produce a scored, synthesized answer.

What PolyGnosis does

PolyGnosis routes a hard problem through a fixed protocol. An orchestrator assigns domain personas such as Security Auditor, DBA Consultant and Backend Architect, and three or more models from different families solve the task independently and in parallel. A dedicated critic model then reviews every solution. The LLM gives per-axis scores from 0 to 10 across five dimensions, and Reciprocal Rank Fusion with Borda Count ranks the results deterministically. A synthesizer merges the strongest parts, and a Constitutional Quality Gate falls back to the best single solution if the merge is worse.

Personas control tool access through hermes chat -t: read-focused personas get web and file, while builders also get terminal. An Early Resolution circuit skips critique and scoring when the solvers agree, and a Reflexion buffer stores severe bugs for later solver prompts. You copy the repository into ~/.hermes/skills/research/polygnosis, set model aliases in config.yaml, validate the config with the bundled script and then ask Hermes to use PolyGnosis. It is licensed GPL-3.0.

Key features

  • Parallel solving by three or more model families from assigned expert personas
  • Dedicated adversarial critic that reviews every solution
  • Deterministic ranking with Reciprocal Rank Fusion and Borda Count
  • Constitutional Quality Gate that rejects a synthesis that regressed
  • Early Resolution shortcut when solvers agree
  • Reflexion corrections buffer for severe bugs

When to use it

  • Designing production code, such as an auth middleware or a connection pool, where a wrong answer is costly
  • Having several models cross-check a security-sensitive design
  • Reducing single-model overconfidence on complex engineering tasks

Who it is for: Hermes Agent users with access to several model providers who want cross-checked answers on high-stakes engineering tasks.

How it fits with Hermes Agent

Built for Hermes Agent as a skill installed under ~/.hermes/skills/research, and it uses hermes chat -t to give each persona its own tool set.

How to install PolyGnosis

These commands are copied from the project's README. Check the repository for the latest steps before you run them.

git clone --depth=1 https://github.com/mosesman831/PolyGnosis.git /tmp/polygnosis
rm -rf /tmp/polygnosis/.git
cp -r /tmp/polygnosis ~/.hermes/skills/research/polygnosis
rm -rf /tmp/polygnosis

Requirements: Hermes Agent, plus model aliases in the skill's config.yaml for the orchestrator, three solvers, critic, synthesizer, meta_reviewer and fallback roles

FAQ

What is PolyGnosis?

PolyGnosis is a Hermes Agent skill that runs an adversarial consensus protocol. Several models solve a problem separately, a critic attacks each solution, ranking algorithms pick the best, and a synthesizer produces the final output.

How do I install PolyGnosis?

Copy the repository into ~/.hermes/skills/research/polygnosis, edit config.yaml with your model aliases, run scripts/validate_config.py, then ask Hermes to use PolyGnosis for a task. The README shows the exact copy commands.

Is PolyGnosis free and open source?

Yes. The repository is licensed under GPL-3.0, a copyleft open-source license.

Similar multi-agent for Hermes Agent

All multi-agent

Related guides: Build agent teams with the Hermes kanban board · Run multiple Hermes agents with profiles · Connect Hermes agents on several machines with Hermes Desktop