Agent Operating System: How to Run Multiple AI Agents

01What Is an Agent Operating System?

An agent operating system is a central place where you plug in all your AI agents, like Hermes Jarvis, and organize them in one system. Think of it as a dashboard that connects your agents, pipelines, and tools so they work together smoothly. In this episode, Julian Goldie shows how he uses it to run content creation, SEO, and even a game studio.

The system makes it easier to manage agents, build custom workflows, and automate tasks like blog posts or videos. Instead of juggling multiple tools, you get everything in one spot. It's like having a command center for all your AI helpers.

Julian explains that this is the future of building with AI. You can see your agents, give them tasks, and watch them complete work automatically. For example, he built a research tool that checks websites for keyword opportunities and deploys content to his site without manual steps.

02Main Features of the Agent OS

The agent OS in the episode includes several key features that help you control your agents. One is the Kanban board, which lets you orchestrate teams of agents. You can give tasks, see progress, and automate projects like blog posts and video edits.

Tasks move from to-do to done with minimal effort.

• Group chat: Agents can chat together in the Agent Mastermind section to solve problems.
• Pipeline section: Agents build and create together in a step-by-step flow.
• Paperclip teams: Use Paperclip to run agents as a team for bigger builds.
• Obsidian integration: Connect Obsidian as a shared memory system for your agents.
• Custom skills: Save skills for tasks like landing pages or emails, so agents reuse them.

Julian also mentions OpenRouter and Fusion, a feature that uses five models to answer the same question and then fuses the best answer. It's a one-shot model that takes five to ten minutes per answer but gives high-quality results for important decisions.

03How to Set Up an Agent Operating System

Setting up your own agent OS might sound techy, but Julian says you don't need to be a coder. He shares how to get started, even if you're a beginner. You can build it with AI help, like using Claude to implement your ideas.

1. Download the agent OS zip file from Julian's classroom (if you have access).
2. Install Obsidian and set up a local vault as your memory system.
3. Tell your agent (like OpenClaw or Hermes) to use Obsidian as memory, giving the file path.
4. Set up your first agent, like Hermes, and connect it to the OS.
5. Add your tools and CLIs, such as Claude or Grok, as needed.
6. Create workflows for your tasks, like SEO or video content.

Julian emphasizes that it's simple enough to set up in 20 to 30 minutes. He mentions community members who got it running in under a day. The key is to focus on one step at a time, like setting up Obsidian first before adding more agents.

04How to Use the Agent OS: Examples from the Episode

In the episode, Julian walks through several examples of using the agent OS. One is creating an SEO keyword research report. He gives the system a task like "create an SEO keyword research report for topics around agent OS," and the system scopes the goal, organizes the research approach, and assigns an actionable to a content writer.

It works smoothly.

Another example is fixing stuck tasks. If a Kanban task gets blocked, you have four options: comment on the task, move it to the to-do list, ask Hermes directly to fix it, or use Claude to adjust the system. Julian shows that asking Hermes to unblock itself works because it's self-learning.

For clients, he suggests building a simplified version. Remove custom workflows and only keep the automations they need, like lead tracking or video creation. You can use Claude to help build those features, test them, and then demo to the client.

He also shows a community member's customization with Comfy UI for local video generation and GLM 5.2. This lets her generate videos and use skills for landing pages or emails. It's a great example of making the OS your own.

05Results and Comparisons from the Episode

Julian tested OpenRouter and Fusion and found it performed well on his GoldieBench. It tended to outperform other models using the same problems. He shows examples like a smooth RPG game created with Fusion.

However, it takes five to ten minutes for each answer, so it's not for everyday coding. He recommends it for big builds or important decisions, then switching back to Claude for regular work.

• OpenRouter/Fusion: Best for final answers, tough decisions, and high-quality outputs.
• Hermes: Good as a main agent, works with free APIs or OAuth logins.
• Claude: Another top pick for the main agent; Anthropic uses Claude for 80% of its code.
• GPT 5.5: Also solid if you're already using it.

Julian also shares that his Agent OS saves time compared to tools like Hermes Workspace, which had sync issues. He built the OS to avoid those problems. For clients, he recommends simplifying to prevent confusion and bugs.

06Tips, Limits, and Mistakes to Avoid

Julian gives practical tips to get the most from your agent OS. First, when setting up output paths in Paperclip, remember that paths in skill files are suggestions, not rules. To enforce paths, change the working directory in the agent's runtime settings.

Use relative paths with the correct CWD.

He also warns about technical bugs, like with models like Olama. If a local model doesn't work, switch to a normal API or a CLI, like Claude or Grok build CLI. You can ask the agent that set up the system to make that change.

• Keep tasks simple and focused—one goal at a time.
• Don't overload clients with unnecessary features; simplify the OS.
• Use Fusion only when you need the best answer, not for daily tasks.
• Set up Obsidian as memory to help agents remember and organize.

Avoid assuming the OS works perfectly. Always test new features yourself and watch for sync issues. Julian stopped using Hermes Workspace because it didn't sync properly after updates, so he recommend the OS for stability.

07Who the Agent OS Is For and Why It Matters

The agent OS is for anyone who wants to get more done with AI, not just programmers. Julian says he's not a coder at all, but he can build complex systems because AI helps him. This makes it accessible for business owners, marketers, and creators who want to automate tasks like SEO, video editing, or lead tracking.

It matters because it centralizes your AI agents, making them work together efficiently. You can run multiple businesses or content pipelines without jumping between tools. It also allows customization—like adding Comfy UI for local video or skills for specific tasks—so the system grows with you.

If you're building for clients, the OS is perfect for delivering simple, customized solutions. You can remove extra features and focus on what the client needs. This saves time and reduces errors, making it a valuable tool for agencies and freelancers.