01What Is an Agent Operating System?

An agent operating system is a central hub where all your AI agents live and work together. It lets you run your whole AI team from your phone, plug in new models as they appear, and automate tasks so you get results even while you sleep.

Think of it like a dream team: every agent has a role, they share the same rules, and they improve every single day you use them. Instead of juggling separate apps and projects, everything slots into one place.

02Key Features of the Agent OS

The agent OS brings several powerful features together:

  • Centralized control: All your agents sit side by side, so you can manage them from one dashboard.
  • Plug-and-play models: When a new useful model drops, you can add it to the system right away. For example, you can set up local models in a local engine or use a free AI coder.
  • Mobile access: Run your entire AI team straight from your phone.
  • Background automation: With tools like Hermes, scheduled tasks keep running even when you close the app or switch off your computer.
  • Transferable skills: Skills are just markdown files, so they work across different agents like Claude, ChatGPT, or Hermes.

These features save time, reduce repetition, and help you get more done with less effort.

03How to Set Up an Agent Operating System

Setting up your agent OS isn't a single install—it's a way of organizing your agents and skills. Here’s how to get started:

  1. Choose your agents: Pick the ones you already use, like Claude, ChatGPT, or Hermes. Hermes is preferred for scheduled tasks because it runs in the background.
  2. Create one central system: Instead of separate OS setups per project, keep one agent OS to avoid multiple points of failure.
  3. Write skills as markdown files: These are simple text files that tell any agent how to behave. You can transfer them between agents.
  4. Set up remote access: Use Tailscale to connect your home computer to your phone or laptop. If that doesn’t work, try a VPS or Cloudflare tunnel.
  5. Decide on local vs. cloud models: For weak computers, skip local models and use APIs or a VPS instead.

Start simple. Add one project at a time. Test, iterate, and refine your instructions until the outputs are exactly what you want.

04Using the Agent OS Day-to-Day

Once your agent OS is running, everyday use becomes about refining instructions and plugging in what you need. For example, if you use Claude or ChatGPT projects to create reports or sales pages, don’t expect perfection on the first try. Push back, give feedback, and ask for multiple options.

For scheduled tasks, use Hermes because it runs in the background even when your computer is off. If you want to use Gemini with a coding agent, try the anti-gravity CLI directly instead of plugging it into Hermes.

For SEO content, you can build a skill that takes a keyword and a case study, automates the writing, and deploys it to your website. The skill includes many rules: keyword quality control, formatting rules, title formulas, and meta description rules. This keeps content quality high.

05Results and Comparisons from the Episode

Julian shared practical experiences from the community. Project-based tools like Claude and ChatGPT often get things wrong on the first try, but usually improve within one or two responses. By giving detailed examples and rules, you can train them over time—sometimes with 30 or more iterations.

He also tested GLM 5.2 and found it “not bad,” but for cheaper open-source options with subscriptions, he recommends Quen 3.8 on the coding plan or KimiK 3. These plug straight into Hermes.

For local models, even a new Mac Studio isn’t great compared to frontier models. So for strong results, use cloud APIs. The system improves every day with use, and you can check the bench to see real builds.

06Tips, Limits, and Mistakes to Avoid

Avoid these common pitfalls:

  • Don’t create multiple OS setups: One agent OS is enough. Multiple setups create many points of failure and overwhelm you.
  • Don’t expect first-try perfection: Even with good projects, the first response may be boring or off. Provide feedback and ask for multiple options.
  • Don’t use local models on old computers: Use APIs or a VPS to offload the heavy lifting.
  • Don’t lock skills to one agent: Skills are transferable as markdown files—test them on different agents.
  • Be patient: You are ahead of most business owners. Training an AI agent is less frustrating than training a human employee for six months.

If you close an app like Claude or codex, scheduled tasks may not run. Hermes is the reliable choice for background tasks.

07Who Should Use an Agent Operating System

Anyone who runs multiple AI agents, creates content, or automates tasks can benefit from an agent OS. It’s especially useful for business owners who want to save time and scale their output without hiring more staff.

If you’re just starting, keep it simple. Use one system, plug in your best skills, and let it improve daily. The system works harder than any employee and runs while you sleep, so you can focus on growing your business.