01What is Hermes Agent Agent OS and how it works

Hermes Agent Agent OS is a system that organizes your AI agents and profiles so you can switch between models and run workflows easily. It lets you set up profiles for different models, like Claude Opus 5.5 or local models, and control them from one place.

In this episode, Julian uses Hermes Agent with Agent OS to answer questions about updating the system, fixing issues, and using skills. He also shows how to automate video editing with HeyGen and Opus 5.5, and how to set up multiple profiles across devices.

02Main features and what's new in Agent OS

Agent OS gives you a drop-down list to choose which model profile to use. You can have profiles for cloud models like Claude Opus 5.5, local models like LFM 2.5 2.6B, and even a Hermes cloud profile.

  • Update Agent OS easily using an update.md file.
  • Create profiles for different models and switch between them.
  • Fix problems by asking Claude desktop to configure and test everything.
  • Link multiple computers with Tailscale to share the same Agent OS.
  • Use skills to save and reuse daily workflows.

Julian says this is really new and only possible in the last week or so, especially for editing videos with Opus 5.5.

03How to update Agent OS step by step

To update your Agent OS, you don't need to reinstall everything. Just get the latest version and ask your AI agent to do the work.

  1. Go to the Agent OS inside the classroom and get the latest version.
  2. Open a new chat in Claude or Codex, wherever you set up Agent OS.
  3. Attach the Agent OS zip file to the chat.
  4. Type: "Can you update the Agent OS using the update.md file?"
  5. Let the agent work in the background and configure everything.

Julian says the update might take a few minutes, but it's automatic.

04How to fix common Hermes Agent OS issues

If something isn't working, you don't need to start over. Instead, go to the agent that set up the system and ask it to fix the problem directly.

For example, if you can't switch between OpenAI, Gemini, and NewsPortal, tell Claude desktop: "Set up three different profiles inside Hermes agent and make sure I can switch between them. Test that they work and fix any issues."

Julian shows that the agent will create profiles, test them, and ask you to log in if needed. If a profile fails, the agent fixes it and tests again. It might take 5 to 10 minutes, but it's step by step.

  • Break down big tasks into small steps, like setting up one model at a time.
  • You don't need every feature installed, only the ones you use daily.
  • Go back and forth with the agent until everything works.

05How to set up multiple Hermes Agent profiles

For a hybrid setup with local and cloud profiles, you just create multiple profiles inside Hermes Agent. Each profile can be for a different model, local or cloud.

Julian says he has profiles for Claude Opus 5.5, LFM 2.5 2.6B (a local model), and Hermes cloud. To organize them, he uses Claude desktop.

  1. Open Claude desktop.
  2. Ask it to add a new profile for a specific model.
  3. Provide the API key or local model details.
  4. Let Claude configure the profile and test it.

For multiple computers, he uses Tailscale to link all devices to the same Agent OS. That takes about 10 minutes to set up. Obsidian is for memory, so it doesn't affect profiles.

06How to build and use skills with Hermes Agent

Skills are for daily repeated workflows. Julian says to save a skill when you do something every day, like SEO or video creation. Skills stay saved, and you can move them between Claude and Codex by sharing the skill.md file.

The process has three steps:

  1. Describe what you want the agent to do.
  2. After it completes the workflow, say "Save that as a skill."
  3. Test and iterate – give feedback to improve the skill.md file.

For example, if your video agent creates a video, you can say "Save this workflow as a skill" so you can reuse it later. To improve it, give feedback and ask the agent to update the skill file.

  • Use skills minimally – don't create one for everything.
  • Too many skills make the agent bloated and confused.
  • Focus on daily workflows only.

07What about local models and technical questions

Almir asked about local models failing on an RTX 5070. Julian explains that he runs local models with Ollama and uses the defaults. He doesn't worry about technical details like thinking mode, ROPE, or YARN.

For constrained output or grammar/schema, Julian says that's technical jargon and not something he worries about. Instead, he recommends using Claude desktop to configure and tweak everything until it works.

For small models, he suggests LFM 2.5 2.6B. It runs fast and smooth even on a Mac Studio, and it's good at tool calls because it was trained with Hermes Agent as a harness.

  • Use Ollama defaults for local models.
  • Try LFM 2.5 2.6B for a lightweight and fast option.
  • Get Claude desktop to handle technical tweaks for you.

08Tips, limits, and mistakes to avoid

One big mistake is trying to set up everything at once. Julian says to break it down into steps, like setting up one model per day. Otherwise it feels overwhelming.

Another mistake is creating too many skills. Only save skills for daily workflows. Too many will bloat the agent and confuse it.

When automating video editing with Opus 5.5, be patient. A 30-second video can take about 10 minutes to edit. A 10-minute video might take 30 to 60 minutes.

Also, don't paste API keys in chat. Put them in the .env file instead.

09Who is Hermes Agent Agent OS for and why it matters

Hermes Agent Agent OS is for anyone who uses AI agents daily and wants a simple way to manage different models and workflows. It's useful for creators, SEOs, and business owners who want to automate repetitive tasks.

Julian says it's powerful because you can update, fix, and expand your system without starting over. You can also link devices with Tailscale, so your agents work together no matter where you are.

The system also helps you learn and build skills that save time. By using it, you can focus on your work instead of technical details.