Hermes Mixture of Agents is ABSURD!

**Julian Goldie** (0:00)
Hermes Mixture of Agents is absolutely wild. So this is a new system where you can combine different models. And the goal here is that you have a panel of agents that work together and get better levels of intelligence than frontier models. So you can see some examples of what we created here. So for example, we actually created a full Windows style operating system. Looks absolutely beautiful. And this was created using Mixture of Agents from Hermes. Now, Hermes itself is just an agent, but what we're doing here is we're mixing different models together. So for example, we have Claude Opus 4.8 and we have GPT 5.5, and then we can choose who is the aggregator. Now, if you're wondering what is an aggregator, that is basically the model that will fuse the answers together. And then from here, we've created this system where we can basically ask the panel anything. So if we type in, for example, a prompt here, we can then run the panel and then work together. And whatever we create goes directly into our workspace. And it's creating some pretty amazing stuff as you can see right here. Now if you're wondering how does this work? Well, the main thing that I would say here is that the model doesn't matter anymore, but the system does. And News Research, who actually created Hermes Agent, basically now allows you to stack several models inside Hermes. And this can be Opus and GPT, not with a bigger or better model, but with a smarter system. So let me explain how this works step by step. And by the way, if you're wondering, okay, like have I tested this? Or do you have a demo, et cetera? Do you have side by side comparisons? So we actually tested 42 different builds with Hermes Mixture of Agents. And again, this is a panel of frontier models merged by a chair, which means that you have a system that works together. Now with this system, we've created loads of different builds. You can actually see it on Goldie Bench. And then what we can actually do is compare this side by side versus something like Claude Opus 4.8. So this itself was built with this Mixture of Agents system. So the point here is like Opus 4.8, amazing model, fantastic. Probably the best one out there. But when you have two models working together, they're always going to outperform one model working alone because two minds are greater than one. So if we compare Opus 4.8 versus Hermes Mixture of Agents, there's actually quite a big difference here. So let's take a look at these two examples. So this is a crypt game that we created.
Looks super nice, great background, etc. And this was created with Mixture of Agents. Now, we actually asked Opus to do the same thing and it failed. We can't get through the start screen, as you can see. Now, it didn't fail on everything. So if we compare these side by side, this is created with Mixture of Agents.
Again, looks super cool, kind of feels like the original game, etc.
Actually a lot of fun to play.
And then if we compare that side by side versus Opus 4.8, it doesn't feel quite as nice. I mean, it's still good, but that's the point, is if you want the best, then you would combine multiple agents. And the point to note here is like, Mixture of Agents is not a model, it's a system.
Whereas Opus 4.8, that model will change. And the thing here to note as well, is that whatever models change and whatever models come out next, so for example, like Sonnet 5 is rumored to come out next, then we can combine that with GPT-5 or GPT-56 when it comes out and get even better outputs. So the whole point is that we get higher levels and better levels of intelligence using the system. Now, when you're running it, as you can see here, the panel will deliberate, and then you just have to wait for the outputs. The thing that I would say here as well, is it's not all perfect because when you're using Mixture of Agents, it can take quite a bit longer to number one, get both answers from both models and then fuse them together to get a good answer. So it can take a little while. And also these would both run on APIs, of course, as well. So Opus 4.8 you can't use with the CLI if you're using Mixture of Agents. So those are two honest disadvantages to just make sure you understand before we go further.
But the point is, if you can get better levels of intelligence, well, that's absolutely awesome. Now, the other thing to note here is that if you're using Mixture of Agents normally, for example, inside the terminal, it's very difficult to manage them and it's quite technical. Let me show you an example of the instructions for using this.