Hermes is a harness — the body around a swappable brain. Understanding its layers is what stops you treating every new model launch as a reason to start over. Here’s how profiles, persona, memory, skills and scheduling actually fit together.
Short answer
- Hermes is a harness, not a model — the brain is a swappable part.
- Profiles are the load-bearing idea: isolated config, memory, skills and history each.
- CLI, desktop, dashboard and messaging gateways are views onto the same agent.
- Set up memory and persona first; interfaces matter least despite being most visible.
Model, harness, and everything else
The single most useful distinction in this whole space: an AI model is a brain, and a harness is the body around it.
A brain on its own can’t open a file, remember yesterday or use a tool. The harness supplies the hands, the memory, the workspace and the rules. When you hear “AI agent”, you’re hearing model plus harness.
Hermes is a harness. That’s why it works with any model you point it at — frontier, free, or local — and why swapping the model doesn’t cost you your setup. The brain is a part; everything around it is Hermes.
Once that clicks, the rest of the architecture makes sense.
The layers
| Layer | What it holds |
|---|---|
| Profiles | The unit of isolation — each one its own config, memory, skills, credentials and chat history, in its own folder |
| Persona (SOUL.md) | Who the agent is: how it talks, what it cares about, what it’s allowed to do |
| Memory | Persistent context across sessions — your business, your decisions, what worked last week |
| Skills | Reusable abilities saved as plain markdown, which the agent writes itself when it solves something hard |
| Tools | What it can actually do — read and write files, run commands, search, browse |
| Scheduling | Cron jobs for unattended work; /loop and /goal for in-session automation |
| Interfaces | CLI, TUI, desktop app, web dashboard, and gateways to Telegram, Discord, Slack, WhatsApp and email |
The thing worth noticing is that these are separable. Your memory isn’t trapped in the model. Your skills aren’t trapped in the interface. Your persona isn’t trapped in the machine. That separation is what makes the whole thing portable.
🔥 Want this set up without the guesswork? Understanding the layers is one thing — knowing which to set up first for a business is where people get stuck. Inside the AI Profit Boardroom you get the Agent OS as a ready-to-install file, a 30-day roadmap, daily tutorials the same day new tools ship, and four live coaching calls a week where you share your screen and get unstuck. → Get access here
Profiles are the load-bearing idea
Almost everything interesting in Hermes is a profile underneath.
A profile is a fully isolated setup: its own config, its own memory, its own skills, its own credentials, its own chat history, stored in its own folder on your machine. Which means:
- You can run several agents that genuinely don’t contaminate each other.
- You can point each one at a different model — a research profile on one, an inbox profile on a small local model.
- You can test two models on the same task side by side and compare honestly.
- A new model release becomes a new profile rather than a migration.
Bot Mode is the clearest illustration of this. It looks like a brand new feature — a roster of named bots with avatars and schedules — but a bot is a profile. Teknium said as much: it’s a new simplified experience for profiles.
That’s why it can’t break the way bolted-on plugins do. It only uses the surface Hermes already exposes.
Same agent, many front doors
The CLI, the TUI, the desktop app, the web dashboard and the messaging gateways are all views onto the same agent — same settings, same memory, same sessions.
You can start something in the desktop app and pick it up in the terminal later. Nothing resets. On messaging platforms the gateway even fires scheduled wakeups between your messages, and the results arrive as ordinary replies.
This is a deliberate architectural choice and it’s the reason the interface question is less important than it looks. Which UI you use changes your experience, not your agent.
How the automation layers differ
| /loop | /goal | cron | |
|---|---|---|---|
| Driven by | A timer, or self-pacing | A judge verdict after each turn | A schedule, outside any session |
| Lives in | Your current session | Your current session | Its own session per run |
| Best for | Polling and monitoring | One objective, iterate until done | Unattended, long-horizon work |
These sit at different levels of the architecture, which is why picking the wrong one is such a common mistake. Sessions are where /loop and /goal live; cron lives outside all sessions. Detail in the loop guide.
The whole agent is portable
The architectural decision with the biggest long-term consequence: your agent’s identity is separate from both the machine and the model.
Export packages the persona, skills, memory, cron jobs, plugins, settings and even desktop layout into a single file — with credentials stripped automatically. Import brings it back on any machine.
Which means models can improve underneath you while you keep the identity. All the tuning stops resetting every time the technology moves. Covered fully in backup and restore.
It also explains why the architecture is worth understanding rather than just following setup steps. Once you know memory, skills and persona are separable from the model, you stop treating a new model launch as a reason to start over.
Where each piece is worth your time
- Memory first. It’s what makes output stop sounding generic, and it’s the piece that can’t be rebuilt from scratch.
- Persona second. It shapes every single response, so it’s the highest leverage per minute spent.
- Skills third, as they emerge from real work rather than being planned up front.
- Scheduling fourth, once you have something worth running unattended.
- Interfaces last. They’re the most visible layer and the least important.
Most people do that list backwards, which is why they end up with a beautiful dashboard in front of an agent that knows nothing about them.
Want the architecture set up rather than explained? The Agent OS in the AI Profit Boardroom ships with the memory layer, profiles, skills and scheduling already wired together, with Hermes running alongside Claude and OpenClaw on one dashboard. Install file, 30-day roadmap, daily tutorials and four coaching calls a week. Start free with the free AI course and community.
FAQ
What is the Hermes agent architecture?
Hermes is a harness — the body around a model. Its layers are profiles, persona, memory, skills, tools, scheduling and interfaces, and each is separable from the others.
What’s the difference between a model and a harness?
The model is the brain that thinks. The harness supplies the hands, memory, workspace and rules. An AI agent is model plus harness.
What is a Hermes profile?
A fully isolated setup with its own config, memory, skills, credentials and chat history in its own folder. Almost everything in Hermes is a profile underneath.
Is Bot Mode a separate system?
No — a bot is a profile. Bot Mode is a new interface layered on the profile system, which is why it uses only the surface Hermes already exposes.
Do the CLI, desktop app and gateways share state?
Yes. Same agent, same settings, same memory, same sessions. Start something in one and pick it up in another.
How do /loop, /goal and cron relate?
They sit at different levels. /loop and /goal live inside a session; cron lives outside all sessions and runs unattended.
Can I change models without losing my setup?
Yes, and it’s the point of the architecture. Memory, skills and persona are separate from the model, so the brain is swappable.
Which layer should I set up first?
Memory, then persona. They shape every response and can’t be recreated later. Interfaces matter least, despite being the most visible.
The bottom line
The Hermes agent architecture is a harness around a swappable brain: profiles for isolation, persona for behaviour, memory for context, skills as markdown the agent writes itself, and several interfaces onto one shared agent. Get memory and persona right first — those are the layers that make everything above them worth having.
About Julian Goldie
I run Goldie Agency, a 7-figure SEO agency, and teach this daily on a 400K+ subscriber YouTube channel. 240+ client projects on Upwork at a 100% job-success score, 10+ years through every major Google update. My systems are in the AI Profit Boardroom; my link building book is free here.
