01What Is Mixture of Agents and How It Works
Mixture of agents is a system that combines several AI models to get smarter results than any single model alone. Instead of waiting for a new frontier model, you create a panel of agents that work together.
In this episode, Julian Goldie explains how Hermes Mixture of Agents works. You pick a few models, choose one as the aggregator, and the system fuses their answers. The goal is to beat top models like Opus 4.8 and GPT-5.5 without needing a bigger or better model.
The key idea is that the model doesn't matter anymore, the system does. Two minds are greater than one, so a panel of models can outperform one alone.
02Main Features of Hermes Mixture of Agents
Hermes Mixture of Agents has several key features that make it powerful.
- Combine multiple models – You can mix models from different providers, like Claude Opus 4.8, GPT-5.5, or open models.
- Aggregator system – One model reads all the answers and writes the final response and runs your tools.
- Private analysis first – Each model works alone before the aggregator fuses the results.
- Save mixes as presets – You can save your favorite model combinations and reuse them.
- Switch mid-chat – Use the command /moa to flip between mixes during a conversation.
- Prompt caching – Keeps the system fast even with multiple models.
These features let you build a system that delivers frontier-level intelligence without needing access to gated models.
03How to Set Up Mixture of Agents
Setting up mixture of agents can be technical, but with the right tool it's simple. Here are the steps shown in the episode:
- Choose a tool that supports mixture of agents, like Hermes Agent OS or the terminal CLI.
- Pick the models you want to use. For example, Claude Opus 4.8 and GPT-5.5.
- Select which model will be the aggregator. This model fuses the answers.
- Save your mix as a preset so you can reuse it later.
- Give the panel a single prompt to start working.
- Wait for the models to do their private analysis, then the aggregator writes the final answer and runs your tools.
If you use the terminal, you need to run many commands and follow several steps. But with a tool like Agent OS, you can just type a prompt and see the results in your workspace.
04How to Use Mixture of Agents with Examples
You can use mixture of agents to create all kinds of things. In the episode, Julian showed a full Windows-style operating system made with this system. It looked beautiful and was created from a simple prompt.
He also created a crypt game using mixture of agents. The game had a great background and felt like the original. When he asked Opus 4.8 to make the same game, it failed and couldn't get past the start screen.
To use it, you just type a prompt in the panel. The models deliberate, then give you an output that goes directly into your workspace. You can preview what you created, save it for later, and even compare it side-by-side with a single model.
You can also mix cheaper models to get high-level outputs without paying for expensive APIs.
05Results and Comparisons from the Test
Julian tested 42 different builds with Hermes Mixture of Agents. He compared the results against Claude Opus 4.8 on GoldieBench.
For the crypt game, mixture of agents created a polished game, while Opus 4.8 failed to get through the start screen. In another comparison, mixture of agents produced a nicer result than Opus 4.8.
Noose Research claims that the system scores 8% higher than Opus 4.8 and 11% higher than GPT-5.5 on benchmarks. Fusion, another system using the same panel engine, even outperforms Claude Fable 5 on benchmarks.
These results show that a system of multiple models can beat a single top model, even when using cheaper or open models.
06Tips, Limits, and Mistakes to Avoid
Mixture of agents is powerful, but it has honest disadvantages. First, it can take longer to get answers because you wait for all models and then the aggregator fusion. Second, the models run on APIs, so you can't use Opus 4.8 with the CLI when using mixture of agents.
Another limit is that managing multiple models in the terminal is very technical and difficult. Julian showed an example of running many terminal commands and following several steps, which isn't accessible to most people.
To avoid mistakes, use a simplified tool like Hermes Agent OS instead of the raw terminal. This lets you see everything you create, preview it, and save it for later.
Also, keep in mind that benchmarks aren't everything. Test the system for yourself and see what works for your use case.
07Why Mixture of Agents Matters Now
Mixture of agents matters because many frontier models are getting gated. Access is granted to only a select few. This means you might not get to use the newest Opus or GPT when they come out.
Instead of waiting, you can build a system using models you already have. Combine open source and cheaper models to get better outputs than an expensive single API.
Three systems have come out recently using this idea: Hermes Mixture of Agents, Fusion, and Sakana Fugu. They all use the same panel engine and get better outputs.
As Julian says, stop chasing models and build the system. This is a proven way to get frontier-level intelligence without needing frontier access.
08Who Should Use Mixture of Agents
Mixture of agents is for anyone who wants better AI outputs without waiting for the next big model. It's especially useful for people who can't get access to gated models or don't want to pay top dollar for expensive APIs.
Non-technical users can benefit from tools like Hermes Agent OS, which makes the system simple and visual. Julian gives this tool away to AR Profit Board members, and the community shares tips and builds.
If you're a developer, you can use the CLI, but be ready for complex commands. For most people, a dedicated dashboard is a smarter way to use it.
In short, if you want to go beyond the public frontier right now, mixture of agents is for you.
