Claude + Hermes Agent: NEW Agent OS is INSANE!

**Julian Goldie** (0:00)
So today, we're going to be looking at how to build your own agent operating system, and some of the best questions I've had about it recently, because I know if those people have questions about this, it's going to help you too, because you probably have similar questions about how to create an amazing agent operating system where your teams of agents can work together, building amazing stuff and just creating awesome stuff. I mean, for example, some amazing things that we've built recently, like for example, what we actually found was that when Fable 5 got taken down, Fusion just came out, right? And Fusion is a new way to have a team of agents working together that can build something better on benchmarks than Fable 5 And I've actually tested this out versus Opus 5, and from what I've seen, it actually outperforms Opus 4.8 by a long way. And so this is really interesting. And what we're going to do today is just answer some of the best questions we have about this stuff, how it works, etc. How to set up Obsidian properly, how to get your Hermes agents working with Claude, how to build best agent operating system possible. So let's get straight into the questions. So one of the first questions we had from Omar is, is it possible for an AI system to take ownership of a single long-term task over an extended period? So for example, like six months, how could you assign it the task of improving a website's SEO performance? So are there any tools or platforms designed for this type of long-running autonomous workflow? So one thing that we have inside our agents is we have the power to build out scheduled tasks. So for example, if you wanted to give Hermes agent the responsibility of organizing and doing something every week for your SEO website, then what you can do is just create a task here, or you can actually go direct to the chat and ask the Hermes agent or whichever agent you use to do this for you. And then you can manage everything in one place if you just go to the Chrome job section inside Manage here. And you can do the same with OpenClaw or any other agent. So how does this work step by step? So you would assign a scheduled task to your agents. Now I would recommend that you either use Claude or Hermes for this for SEO optimization. You would tell it the workflow and tell it every time you do this workflow, you improve it and you look at what's working and what's not working, and you iterate based on that and you update your workflow. And then also run it on a daily schedule. So for example, it could be daily optimization of meta-tasks. It could be uploading one new blog per day.
And that's how you can give your agents responsibility for one task, because it's running on a scheduled system. And then also you've got the manage section here with scheduled tasks, so you can easily see or delete any stuff that you need.
And to assign a task, you just go to the chat with your agent and assign it directly there. So Andrea has a question here, which is Hermes is failing to use Claude code for coding work consistently.
So when it actually comes to coding work, I typically prefer to use Claude code in the first place. So for example, actually like building out the agent operating system on the back end, we'll use Claude desktop because it just tends to do coding work flows better. Whereas Hermes is great for just doing little tasks or scheduled tasks or recurring tasks. Now I'd also question which model you're plugging into that too.
So there's two aspects to this. Number one, I would actually use Claude for coding work instead of Hermes, simply because Hermes is great for scheduled tasks or admin tasks or smaller tasks, but for coding I tend to find I get better results with Claude. So when I'm building out my agent operating system, we actually use Claude desktop on the back end. And then when it comes to following instructions and skills properly, I'd also look at which API you've plugged into Hermes. And if it's not doing what you want, I would look at using something else instead. So for example, you could switch to Minimax M3, you could switch to Grok, et cetera. Quite often, if Hermes is not working that well, it can be down to the brain that's plugged into the agent as well. So Carl was posted about his current build for his marketing agency.
Basically, he's got a current setup with he's got like Javis as dispatcher. Subagents are created from there with a team of advisors and specialists. And he's got a knowledge base with Obsidian now too. But the hard part he's saying is not can AI do the task. You know, he's pretty good at that. But the problem is getting reliable handoffs, clean source rules and enough QA on it as well. So how can you build better systems throughout quality control basically? So this is actually something that we've been testing out lately to get better results and it's working really nicely. So if we go over to the Kanban board here, I typically in big tasks, like let's say for example, we're trying to create a fully edited video with AI and just, you know, having minimal input, but also making sure that everything that you do is set up nicely together.