Most people meet Hermes the same way I did: a chat window on a screen. But the Hermes agent architecture sitting underneath that window is one of the smarter designs in AI right now, and once you understand how it is built, you stop treating it like a chatbot and start treating it like a member of staff.
📺 Watch: Hermes Agent OS Just Changed AI Agents Forever!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
I run Hermes every day across my SEO business, my YouTube channel and my community. I have pulled it apart layer by layer, swapped its brain out mid-week, moved it between machines and watched it teach itself new abilities overnight. This is the full teardown: the nine layers, what lives in each one, and why the design decisions matter far more than any single feature.
The mental model in one sentence
Hermes is a free, open-source, self-improving agent framework built by Nous Research and released under an MIT licence. The mental model that makes the entire Hermes agent architecture click is this: the harness is the body, and the model is the brain. Everything that makes your agent yours — identity, memory, skills, schedules — lives in the body. The brain is a replaceable part.
Hold onto that sentence. Every layer below makes sense because of it.
Layer one: the brain
Hermes is provider-agnostic by design. The model is literally a profile setting, not a foundation. You can run Claude one week, DeepSeek the next, then Grok, GLM or a fully local model, and the agent stays the same agent. Each model gets its own profile with separate histories, so nothing bleeds between brains.
This is the layer most people get wrong. They think the model is the agent. It is not. The agent's identity is not the model — the model is just the reasoning engine currently plugged into the socket. I decide which brain wins with side-by-side testing on Goldie Bench, and this architecture is exactly why that testing pays off: when a new model tops the board, swapping it in costs me nothing. No migration. No retraining. Same agent, sharper brain.
Layer two: the agent loop
A chatbot answers. An agent acts. The loop layer is what separates the two.
Hermes reads files, writes files, runs commands and searches the web. It ships with 30 native tools, and it uses them inside a proper working loop: plan, act, check the result, repeat. Ask a chatbot to fix your pricing page and you get advice. Ask Hermes and it opens the file, makes the edit, checks the output and reports back on what changed. That loop running against real tools is the difference between talking about work and doing it.
📺 Watch: Hermes AI Agents Just Went Portable
Layers three and four: skills and memory
The skills layer teaches itself
Hermes ships with 80 skills, and each one is nothing more exotic than a plain markdown file: a name, a when-to-use description, and step-by-step instructions. That plainness is deliberate. You can read every ability your agent has, edit any of them, and write your own in minutes.
The clever bit is the skill loop. When Hermes solves something genuinely hard, it saves the method as a new skill file — self-taught abilities that persist. That mechanic is exactly what sits behind my PDF book-brain setup, where a working method became a permanent skill rather than a one-off win. Type /journey and you get every skill and memory laid out on a timeline, so you can literally watch your agent get more capable week by week.
If you want this architecture working for your business without building it alone, the AI Profit Boardroom ships the Agent OS with Hermes wired in properly. → Get the working build
The memory layer never resets
Hermes keeps persistent memory across sessions. It remembers you, your business, your preferences and the lessons from past work. In my setup it pairs with an Obsidian vault, which I covered in my guide to the best memory for a Hermes agent.
The architectural decision that matters: memory lives in the agent, not the model. Swap brains and nothing is forgotten. Your Claude-powered agent and your DeepSeek-powered agent share one accumulated understanding of who you are, because that understanding was never inside either model.
Layers five and six: surfaces and profiles
One agent, many faces
Hermes is one agent with many faces: a CLI, a TUI, a desktop app for Mac, Windows and Linux, a web dashboard, and a gateway out to Telegram, Discord, Slack, WhatsApp and email. The desktop app even has a preview rail that renders pages and files beside the chat — I broke it down in my Hermes desktop browser review.
Same settings, same memory, same sessions everywhere. Start a job in the desktop app, continue it in the terminal, nudge it from Telegram while you are out — nothing resets, because every surface is a window onto the same agent rather than a separate product. If you are deciding where to run all this, I have also tested the best OS for a Hermes agent.
Profiles, and why Bot Mode is simpler than it looks
Each profile is an isolated folder on your machine with its own config, memory, skills, credentials and chat history. That isolation is what lets you run a researcher, a writer and an operations agent side by side without them contaminating each other.
Bot Mode is just a friendly interface laid over that system: a roster of named bots, each with its own model, memory and schedule. Teknium called it a new simplified experience for profiles, and that is the right way to read it — nothing new underneath, just the existing profile architecture made visible.
Layer seven: automation
This layer turns Hermes from a tool you use into a system that runs. Cron jobs handle scheduled work outside your sessions. /loop runs recurring turns inside a session, on a fixed interval or self-paced. /goal sets a judge-driven objective the agent keeps working toward. Add session heartbeats, subagent delegation and a multi-agent Kanban board, and you have genuine orchestration rather than one clever assistant.
The docs also describe a Codex app-server runtime integration, and the release cadence keeps pushing this whole layer forward — I covered the latest wave of changes in my Hermes v0.20 breakdown.
📺 Watch: Grok Bot DESTROYS Hermes Agent?
Layer eight: interop
Hermes speaks the A2A protocol v1.0, which means it can discover other agents, talk to them, and even be driven by them. It also reads the shared instruction files other harnesses use — agents.md and claude.md — so it slots into an existing setup instead of demanding you rebuild around it. For research work it produces grounded citations and has a fact-checking mode, which matters enormously when the output is going anywhere near a client.
Layer nine: portability
The whole identity — persona, skills, memory, cron jobs, plugins — packages into a single export file with your keys stripped out. Import that file on another machine and your agent is simply there, intact. I walked through the process in my guide to Hermes agent backup and restore. The principle underneath it: your agent's identity is separate from any machine, exactly as it is separate from any model.
Why the Hermes agent architecture actually matters
Pull the nine layers together and the thesis is simple. Harness equals body; model equals brain. Because identity, memory and skills live in the harness, every model upgrade is a brain transplant that keeps the person. The industry ships a better brain every few months, and your agent absorbs each one without losing a day of accumulated context.
And because everything is files on your machine — markdown skills, memory, profile folders — you own all of it and can move all of it. No vendor holds your agent hostage. That combination of compounding capability and genuine ownership is why I think this design ages better than anything else I have tested.
The full stack at a glance
Here is the entire Hermes agent architecture on one screen.
| Layer | What lives there | Why it matters |
|---|---|---|
| Brain | The current model — Claude, DeepSeek, Grok, GLM or local — one profile per model | Swappable reasoning; identity never trapped in a vendor |
| Agent loop | 30 native tools; plan, act, check, repeat | Turns answers into finished work |
| Skills | 80 markdown skill files plus self-saved methods | Abilities you can read, edit and grow |
| Memory | Persistent memory across sessions | Context survives every brain swap |
| Surfaces | CLI, TUI, desktop app, web dashboard, messaging gateway | One agent everywhere; nothing resets |
| Profiles | Isolated folders: config, memory, skills, credentials, history | Run a team of agents without crossover |
| Automation | Cron, /loop, /goal, heartbeats, subagents, Kanban board | Work happens while you sleep |
| Interop | A2A v1.0, shared instruction files, grounded citations | Plays well with other agents and harnesses |
| Portability | One-file export of the whole identity, keys stripped | Your agent outlives any machine |
What this unlocks in practice
None of this is theory for me. I run named profiles — Oracle for research and strategy, Astros for operations — each an isolated folder with its own brain and its own job. Memory pairs with my Obsidian vault, so every session starts already knowing the business. Cron owns my mornings: the overnight work is waiting before I have made coffee. The whole stack is the backbone of my Agent OS, which is really just this architecture used deliberately instead of accidentally.
Hermes agent architecture FAQ
Is Hermes open source?
Yes. It is free and open source under an MIT licence, built by Nous Research. You can read the code, run it on your own machine and modify anything you like.
Does the model matter?
Yes — a sharper brain gives you a sharper agent. But the architecture makes the model a swappable setting rather than a commitment, so the real question is not which model forever, it is which model this month.
Where does memory actually live?
In the agent, as files on your machine — not inside the model. That is why you can change providers and your agent still remembers your business, your projects and every past lesson.
Can Hermes talk to other agents?
Yes. It supports the A2A protocol v1.0, so it can discover other agents, communicate with them and be driven by them, and it reads shared instruction files like agents.md and claude.md.
Is my agent portable?
Completely. The persona, skills, memory, cron jobs and plugins export into a single file with credentials stripped, and import cleanly on another machine.
My verdict
The Hermes agent architecture is the first design I have used where the agent genuinely compounds: skills accumulate, memory persists, brains upgrade, and nothing you build is ever trapped in someone else's product. Learn the layers once and every feature stops being a trick and starts being obvious.
If you want an agent that compounds instead of resetting every session, check out the AI Profit Boardroom — inside you get step-by-step Hermes tutorials across every layer of this stack, the Agent OS ready to run, four weekly coaching calls, and 3,700+ business owners building alongside you. → Build on the full architecture
