The Hermes Council Engine.

What the Council Engine is.

It's a new tab I built inside Hermes, in my Agent OS.

Here's the plain version.

You pick a few frontier models — any providers, mixed together. I'm running Claude Opus 4.8 and GPT-5.5.

When you ask a question, both models answer it privately, at the same time.

Then a third model — the chair — reads both answers, judges them, and writes one final answer that's better than either.

The panel of models is the council. The chair is the aggregator. One question in, one better answer out.

Nous Research shipped this idea as Mixture of Agents 2.0. I wired it into a tab, gave it a workspace, and put it on my own leaderboard to see if it actually holds up.

It does. It ranked second on my whole board — above every single model on the market.

A panel of experts beats one genius.

Picture one brilliant person answering a hard question alone.

Now picture a panel — each expert writes their own take, privately — and a sharp chair reads all of them and gives you the best combined answer.

The panel wins. Almost every time.

That's the whole engine. The reference models are the panel. The aggregator is the chair.

One prompt goes in. Several models think. One clean answer comes out — better than any single one of them could give.

It ranked #2 on my whole board.

I didn't want to take anyone's word for it. So I ran the Council Engine through GoldieBench — my one-shot leaderboard.

The test is simple. Ask the model for a real thing — a playable game, a working simulation, a page — in one prompt. Whatever ships on the first try is what gets scored, 0 to 10.

The council built all 42 tasks. Every one. Then each build was judged against the same field every other model faces.

It finished second on the entire leaderboard, with an 8.38 average and 12 task golds.

Second out of every model I've ever tested. Above Opus 4.8. Above Grok. Above every single model in the field.

The only thing ahead of it is Fusion — which is also a system. A different panel of models.

Read that line again as a builder, not a fan: the two best things on my board are both systems. Not models.

What the council actually built.

These are real. Built by the panel, merged by the chair, one prompt each. Touch them — they're live.

Heads up: these are live embeds — the cursor stays free so you can scroll. Hit "play fullscreen" for full mouse-look. Every demo is at goldiebench.com/models/moa.

Where it wins, and where it doesn't.

I'm not going to pretend the council wins everything. It doesn't. Here's the straight version on the big builds.

On the tight stuff, it edges the field. Neon Breakout scored 8.6 — past both Opus 4.8 and Fusion at 8.5.

On the heaviest 3D builds, it trails. The Dragon Realm and the Crypt both landed at 7.8 — while Fusion hit 9.0 on each.

So the council isn't magic. It's a system, and Fusion is a better-tuned system for the hardest builds.

But step back. Both of them — the thing that wins, and the thing in second — are panels of models. Neither is a single model.

That's the real result. You don't climb this board with a bigger model. You climb it with a better system.

Stop chasing the model. Build the system.

Everyone is waiting on the next model. The next Opus. The next GPT.

But look at what just happened on my own board: a mix of today's models beat every single model — and a second mix beat it.

No new release. No gated access. Just a smarter way to use what's already here.

The model is a part you can swap. The system is the thing you own.

And the timing makes this huge. Fable 5 is in preview. GPT-5.6 is delayed. The next frontier models aren't really in your hands yet.

So the winning move isn't waiting. It's squeezing more out of the models you already have — by combining them.

One tab. A workspace for everything it makes.

Inside Hermes there's now a Mixture tab.

At the top it shows the council — your panel models and the chair — live.

You type a prompt, hit run, and watch both models answer in parallel before the chair merges them. You see the final answer and, if you want, each model's private draft.

Under it sits a workspace — every build and every answer the council has ever made, in one place, click to preview. Nothing gets lost.

And it's all configurable. Swap any model into a slot. Save different councils for different jobs. Flip between them with one command.

Why this changes your move.

What's holding you back.

Where this leaves you.

If you only ever use one model at a time, this is the gap you've been feeling.

You're not behind because you picked the wrong model. You're behind because you're using models one at a time, while the systems are pulling ahead.

The people who figure out model systems now — while the tools are this raw — are going to be way ahead when everything settles. Every council you build, every workflow you wire in, it all compounds.

The model expires. The system compounds. That's the whole bet.

The whole thing in five lines.