Hermes MoA: The Model Council Profiles

June 26, 2026 — by James Meadlock

Hermes Agent can already switch between providers and models during a session. The MoA feature — Mixture of Agents — adds a second layer: instead of asking one model for an answer, Hermes can ask one or more reference models for analysis, then send those reference outputs to an aggregator model that writes the final response.

In plain English: MoA turns a Hermes turn into a small model council.

The important part: the reference models are not the final voice. The aggregator is. The references supply competing analysis; the aggregator judges and synthesizes.

The flow

Hermes MoA: model council routing One user turn fans out to reference models, then Claude Opus judges and writes the final answer. Hermes turn /moa council fan-out prompt GPT-5.5 openai-codex reference execution • coding • structure Grok 4.3 xAI OAuth reference independent • contrarian read DeepSeek V4 Pro Fireworks reference third reasoning path Claude Opus 4.8 aggregator • judge • final voice temp 0.25 • max 4096 one synthesized Hermes reply reference models produce competing analysis aggregator reads references, then decides default preset: council | refs temp 0.5 | agg temp 0.25

How to use it

Inside a Hermes session, the command is:

/moa <preset>

The current default preset is council, so a bare command also works:

/moa

To turn MoA off, switch back to a normal model provider:

/model gpt-5.5 --provider openai-codex

The current profiles

PresetReferencesAggregatorUse case
default MoA
council		GPT-5.5, Grok 4.3, DeepSeek V4 ProClaude Opus 4.8Best general-purpose council. Use for strategic decisions, writing, architecture, hard debugging, and “what am I missing?” reviews.
subscriptionGrok 4.3GPT-5.5Cheap second opinion. One reference model, one aggregator.
heavier
deep		Grok 4.3, GPT-5.5Claude Opus 4.7Stronger synthesis than subscription, but lighter than the full council.
defaultGPT-5.5, DeepSeek V4 Pro via OpenRouterClaude Opus 4.8 via OpenRouterOlder profile retained in config. Not the active default.

The council preset

This is the main setup right now:

reference_models:
  - provider: openai-codex
    model: gpt-5.5
  - provider: xai-oauth
    model: grok-4.3
  - provider: fireworks
    model: accounts/fireworks/models/deepseek-v4-pro

aggregator:
  provider: anthropic
  model: claude-opus-4-8

reference_temperature: 0.5
aggregator_temperature: 0.25
max_tokens: 4096

The design is intentional:

I prefer Claude as the aggregator here because the final step should not be a vote counter. It should be judgment: which reference is right, what did each model miss, and what should the final answer actually say?

When MoA is worth the cost

MoA is not the default for every task. It is slower and more expensive because one user turn can become several model calls. That is worth it when disagreement is useful:

It is usually not worth it for routine shell work, simple factual lookup, or small edits. For those, a normal single-model Hermes session is faster and cleaner.

The practical rule

Use normal Hermes for ordinary work.

Use /moa council when the answer deserves a council.

Use /moa subscription when you just want a cheaper second opinion.

More models are not automatically better. Three references plus one aggregator is the current sweet spot. Past that, latency rises, cost rises, and the answer can get mushy. The point is not to build a chorus. The point is to get a few genuinely different reads, then have a strong model synthesize them.

Configuration snapshot from the local Hermes default profile on June 26, 2026. Provider availability and model names may change as subscriptions, APIs, and local routing evolve.

— James Meadlock, June 26, 2026 · al-engr.com