Hermes just released Hermes Mixture of Agents (MoA) — a way to run several AI models in parallel and aggregate them into one stronger answer. Instead of betting on a single model, you build a panel: multiple models think, and a 'chair' model combines them into the best response. It's the same idea behind systems like Fusion and Sakana Fugu, and it's a clever workaround for today's gated frontier models.
📺 Watch: Hermes Mixture of Agents explained + tested
Want Hermes Mixture of Agents wired into a ready-made Agent OS? It's inside the AI Profit Boardroom (with Fusion and Sakana too). → Join AIPB
What Is Hermes Mixture of Agents?
Hermes Mixture of Agents is a virtual model provider. You create named MoA presets, and each one appears as a selectable model under the moa provider. A preset has two parts: reference models that run first and quietly write their own take on the task, and an aggregator — the acting model that reads those takes and writes the final answer (and runs your tool calls). Think of it as a panel of experts plus a sharp chair: one question goes in, several models think privately, and one clean answer comes out that's better than any single one of them.
Why It Matters: Beat The Gated Frontier Models
Here's the timing. The newest top models are gated — Claude Fable 5 is rolling out to a handful of approved partners, and GPT-5.6 is a limited preview — so getting frontier-level intelligence is hard right now. Mixture of Agents is the workaround: combine the models you already have into one virtual model that punches above any single one of them. The lesson is simple — stop chasing the next model; build the system instead. The model is a part you can swap; the system is the thing you own.
Does It Actually Work? The Benchmarks
Yes, and there's data. On HermesBench, a two-model preset — Claude Opus 4.8 aggregating over a GPT-5.5 reference — scores 0.8202, versus 0.7607 for Opus 4.8 alone and 0.7412 for GPT-5.5 alone. That's the panel beating its strongest single member by around 6 points — proof that aggregating a second perspective genuinely lifts quality on hard tasks, rather than just averaging. And because MoA is provider-agnostic, you can mix any providers you like.
📺 Watch: New FREE Hermes Computer Use Agents!
How To Enable And Use Hermes Mixture of Agents
It's a few simple steps:
- Update Hermes first — run the Hermes update in your terminal, or use the Manage section of the dashboard.
- Pick a preset — run
hermes modeland you'll see a Mixture of Agents provider row with your presets. Or select one anywhere with/model default --provider moa. - One-shot it —
/moa your prompt hereruns a single turn through the default MoA preset, then puts you back on your previous model. - Configure presets — from the dashboard (Models → Model Settings → Mixture of Agents), the desktop app,
hermes moa configure, or directly inconfig.yaml.
MoA also works on agent loops, and it composes with goal mode and gateway/desktop sessions because it's just a normal model selection.
The Best MoA Combos (And A Config Example)
The top performer right now is an Opus 4.8 + GPT-5.5 aggregator. A solid default preset looks like this: reference models GPT-5.5 and DeepSeek V4 Pro, with Claude Opus 4.8 as the aggregator. The clever bit — you can even use cheaper models together and still beat a single pricier model working alone, so you hit frontier-quality output for a fraction of the cost. One quick note: each MoA turn makes extra reference calls, so it uses more tokens — you're paying for multiple perspectives, not for broken prompt caching (Hermes keeps the cache intact).
📺 Watch: This NEW Hermes Update Builds Agents From Scratch
Fusion, Sakana Fugu And The Bigger Pattern
MoA isn't a one-off trick — it's a pattern. Fusion and Sakana Fugu run on the same 'panel of models fused into one answer' idea, and both have hit Fable-5-level intelligence in testing. You don't have to rely on just one of them; the smart move is to have all three available and pick the right one per task.
Get It Running In Your Agent OS
I run Mixture of Agents, Fusion and Sakana Fugu all wired into one dashboard — the Agent Operating System inside the AI Profit Boardroom. You get the full Agent OS zip with Hermes MoA, Fusion and Sakana set up, the prompts and presets, coaching calls where we build model panels live, daily tutorials, and a room of 3,600+ operators running this exact stack. → Join AIPB.
Frequently Asked Questions
What is Hermes Mixture of Agents?
A virtual model provider that runs several models in parallel — reference models give their take, an aggregator combines them into one final answer — so you get a stronger result than any single model.
How do I enable Hermes Mixture of Agents?
Update Hermes, then run hermes model and pick a MoA preset, or use /model default --provider moa. Configure presets in the dashboard, desktop app, or config.yaml.
Does MoA cost more?
It uses more tokens because of the extra reference calls, but it doesn't break prompt caching — and it can beat a pricier single model using cheaper ones combined.
What's the best MoA combo?
An Opus 4.8 aggregator over a GPT-5.5 reference tops HermesBench (0.8202 vs 0.7607 for Opus alone).
The Bottom Line
Hermes Mixture of Agents lets you build a panel of models that beats any single gated frontier model — provider-agnostic, benchmark-proven, and just a command to switch on. Stop waiting for the next model; combine the ones you have. And if you want MoA, Fusion and Sakana ready to go, that's the Agent OS inside AIPB.
