01What is Hermes Mixture of Agents?

Hermes Mixture of Agents is a system where multiple AI models work together as a council to answer questions. Instead of relying on one model, it combines the strengths of several models to produce a better final answer.

Here’s how it works: you pick a few frontier models, like Opus 4.8 and GPT 5.5. Each model answers your question privately at the same time. Then a third model, called the chair, reads all the answers, judges them, and writes one final answer that is better than any individual one.

This idea comes from research called mixture of agents. It is built into the Hermes Agent inside the Agent OS. The system is a panel of experts that beats a single genius.

A panel of experts will win almost every time.

You can use it for any question, from coding to creative tasks. The system is simple: one question in, one better answer out.

02Key Features of Hermes Mixture of Agents

The system has a new tab called the Council engine inside the Agent OS. Here are the main features:

  • Pick any providers or models to mix together, like Opus 4.8 and GPT 5.5.
  • Both models answer your question at the same time, privately.
  • A chair model reads both answers and fuses them into a final answer.
  • You get a workspace where you can see all your created outputs.
  • You can switch models whenever you want, making them swappable parts.
  • It runs cheaply because you can use cheap or free models if needed.

You can run the panel with one click. There’s no need to paste answers between different tabs. Everything is in one place.

03How to Set Up Hermes Mixture of Agents

Setting up Hermes Mixture of Agents is easy if you have the Agent OS. Julian shows it as a built-in tab. Follow these steps to get started:

  1. Open the Agent OS and go to the Hermes Agent tab.
  2. Click on the new Council engine tab.
  3. Choose the models you want in your panel, like Opus 4.8 and GPT 5.5.
  4. Type your question in the input box.
  5. Click “Run the panel” to get your answer.
  6. View the results in the workspace side panel.

You don’t need any coding skills. The system is one click away. You can change the models anytime, so you are not stuck with one provider.

04How to Use It: Examples from the Episode

In the episode, Julian used Hermes Mixture of Agents to create games and simulations. He showed a game called Dragon Realm, an open-source project. With one prompt, the system produced a nice 2D-looking game that worked out of the box.

He also made a 3D version of Skyrim-like open world game. That one also looked amazing for a single shot. You give one prompt, you get two answers fused out, and the output is solid.

He compared side-by-side with other tools. For example, the output from Fusion was a cool game, and Hermes’ version was more 2D but still nice. Against Opus 4.8, Hermes produced a neon city simulation that felt less linear and more enjoyable to move around in.

The racing game from Hermes was easy to navigate, while Opus 4.8’s version was buggy and had bad graphics. The system combines the best of both models to give you better outputs.

05Performance on Goldie Bench and Comparisons

Julian tested Hermes Mixture of Agents on his 42-task leaderboard, called Goldie Bench. It ranked number two on the entire board, above almost every single model. The only thing beating it is Fusion.

It outperformed Opus 4.8 by a long way. The results were clear in the side-by-side comparisons. The mixture agent produced less buggy, more useful, and more fun outputs.

Here are the key numbers: It is number two on the leaderboard. It beats Opus 4.8. The top two performers on the leaderboard are systems, not individual models.

That shows that systems give the edge, not just picking a single model.

You can check the results yourself on Goldie Bench. The outputs are hosted right there, so you can see what you would get.

06Tips, Limits, and Mistakes to Avoid

One big tip from Julian is to stop chasing the model. Build the system instead. Everyone waits for the next model or gated access, but you can already get better outputs by mixing today’s models.

Watch out for these limits and mistakes:

  • Don’t wait for models like GPT 5.6 or Fable 5 because they have no confirmed release date.
  • Don’t use multiple tabs to compare models; it’s slower and messier.
  • Don’t think settings are too complex – the Agent OS makes it simple.
  • Don’t ignore cheap models; you can run the system with free or low-cost models.
  • Don’t get stuck with one model; swap them when a better or cheaper one appears.

The system is the thing you own. Models change, but your system stays.

07Who Should Use Hermes Mixture of Agents?

Hermes Mixture of Agents is for anyone who wants better outputs without waiting for the next big model. If you are tired of hitting the ceiling of a single model, this system helps you go beyond that.

It is also good for non-technical people. Julian says he is non-technical, and he could set it up. There are 191 pages of wins and testimonies from users of the Agent OS, so many people are using it successfully.

You can use it for automation, like creating videos or SEO content. Those systems don’t even need a frontier model. They just need good systems.

If you want to automate tasks and get better results, this is for you. You can start with the models you already have access to.

08Why Systems Beat Models

The big lesson from the episode is that systems are more important than models. The top two results on Goldie Bench are systems, not individual models. That proves you don’t need the best model to win.

Most people pick one model and hope it’s the best. But they hit a ceiling and wait months for the next release. With a mixture of agents, you get the best models today and combine them into a virtual council.

You beat every single model on its own. You don’t need extra access. You can swap models when cheaper or better ones come out.

Remember, the models will change in three to six months, but the engine you use to run the system stays the same. So focus on building systems, not chasing models.