15 Hermes Agent Concepts Explained In Plain English

By the end you will know what every part of a Hermes setup does and how the pieces fit together. This article takes you through each one in order.

This article is based on a YouTube video I published walking through every essential Hermes Agent concept in plain English. You can watch it here: https://youtu.be/lGtBPrSrnjY

Here's what you're going to learn

• How the harness and the model work as separate layers, so you can swap the model without rebuilding anything
• How to send different tasks to different models instead of paying for one model everywhere
• How to choose between the terminal, desktop app, Telegram and voice, depending on where you want to work
• How SOUL.md, AGENTS.md, CONTEXT.md, VOICE.md and corrections files decide how your agent behaves
• How to stop a long session going stale with compression and handoff documents, so your agent keeps its working context instead of starting over

1. The harness and the model are separate layers

Think of Hermes like a car. The model is the engine.

Everything else that makes the car move, the wheels, the frame, the brakes, the electronics, that is the harness. The model decides what to say and do. The harness carries your request in, loads the tools and rules, and brings the work back out.

Because those two layers are separate, you can change the brain without changing the machine. I can swap from one model to another without rebuilding my setup.

The tools, the memory, the rules and the sessions all stay exactly where they are.

This matters because Hermes is open source and free, so you can customize the harness to your own use case. I removed a bunch of preloaded tools I was never going to use.

You can add your own rules, add hooks, and shape it around what you actually do.

And once the harness and the model are separate in your head, you can start routing different tasks to different models.

2. You can route different tasks to different models

Hermes lets you treat model choice as a per-task decision instead of a subscription decision.

When I create content, I plan it with GPT 5.6 through OpenRouter. Then I write it with Kimi K3.

I set that up in my agent's file so it happens automatically, with the planning model building the structure and the writing model handling the copy.

3. You talk to the same agent from five different places

The terminal is the classic way. You type your prompt, the answer comes back.

But if you are not comfortable in the terminal, the desktop app gives you the same agent with a different face. It can show you a live preview of what you built, like a landing page or a presentation, right in the window.

I also have Hermes on Telegram, so I can message it from my phone when I am out. The desktop app has a voice mode where I can speak to it and it speaks back.

And I helped a client set up Discord voice last week, so the agent sits inside a voice channel and you can talk to it like a person on a call.

The interface is just where the prompt goes in and the answer comes back. Underneath, it is the same setup every time.

But it still needs to know how you want it to behave. That is what the files are for.

4. The files that tell it who to be and how to work

Without context, your agent is a chatbot. It starts from scratch every time.

The files you give it are what turn it into something useful.

The prompt is the one-off request you are making right now. I dictate mine with Wispr Flow, explaining the project and pointing to files already on my computer.

SOUL.md holds the working style and tone. It tells Hermes what role to play, how direct to be and how to make decisions.

AGENTS.md holds the project rules for a specific folder. CONTEXT.md and VOICE.md hold workspace specifics, like how I want copy written in this particular project.

If Hermes keeps making the same mistake, you save that correction. I put mine in a corrections.md file.

Either way, the mistake gets written down so it does not repeat.

5. The context window fills, so you compress or hand off

Every model has a token limit. Claude Code typically gives you a million tokens.

GPT 5.6 has a 272,000-token context window in Codex, and for me that fills up fast.

The terminal shows a live bar of how full the window is. When it gets to 50 or 60 percent, I stop and write a handoff document.

Then I clear the session, start a fresh one and load that handoff document so we can continue. If I am using an Anthropic model, I do it earlier, around 40 percent.

Hermes can also compress automatically. When you hit the limit, it summarizes the conversation, takes notes on the important parts, and inserts that into a fresh session.

You can set the limit where that happens.

6. The agent loop: prompt, tools, answer

You put in your prompt. Hermes adds it to the context window, then decides what needs to happen next.

It might run a web search, check files in your folder or call an MCP. It takes those actions, then gives you the response.

When I build something, the loop has more structure. The planning phase involves searching files, doing research and coming up with a plan.

Then the build phase begins, where Hermes asks questions to finalize that plan.

Once you are building, Hermes can run an autonomous loop and keep checking the work against the goal until it is done. You do not have to keep prompting it.

7. Tools, MCP servers and plugins

A tool gives Hermes an action it can take outside the model, like searching the web, running a command or calling a service.

I write my newsletter inside Hermes. It has tools that search the web, pull posts from my YouTube channel and Instagram, and then push the finished draft to Beehiiv.

I do not copy and paste anything. There is an MCP for Beehiiv that turns each API endpoint into a tool Hermes can call.

That is what an MCP server does. It wraps an API and makes each endpoint a callable tool.

Think of it like a toolbox. You pick the wrench or the screwdriver depending on the job.

Plugins bundle skills, MCPs, hooks and slash commands into one install. I have the Superpowers plugin installed, which gives me 11 different skills with one click.

Each skill has a specific role and they work together.

8. Memory stores what to remember, skills store how to do it

Memory answers the question of what Hermes should remember, including your preferences, previous decisions and facts about your business.

When you start a session, it checks your memory file and carries that into the context window alongside your prompt.

A skill answers a different question: how Hermes should do a specific job. It is a repeatable method, the steps, rules and resources to use when that type of work comes up.

You do not have to write every skill yourself. You can install skills other people have shared, and you can create your own by typing /learn, which turns a process you just went through into a skill.

Neither one changes the model. Hermes loads the relevant memory or skill into the working context when it is needed.

9. Hooks fire actions automatically when something happens

A hook runs when a named event fires. That could be before a tool is used, after a job finishes, when a new session starts, or when you write a file to your system.

The hook checks your rules in code and either blocks the action or lets it run. It does not take up context.

It happens in a shell, not as a prompt injected into the conversation.

I have a hook set up so my agents do not read my .env file. They can run terminal commands to pull API keys when they need them, but the keys never appear in the chat.

That rule fires on its own, without me approving anything.

Hooks cover the moments inside one setup. Profiles let you run more than one setup at all.

10. Profiles are durable agents, subagents are temporary helpers

A profile is a full Hermes setup built for a particular role or type of work. It has its own SOUL.md, AGENTS.md, memory, skills, models and conversation history.

I have an SEO writer profile, an SEO editor and an SEO strategist. Each one behaves differently because each one has its own instructions and context.

A subagent is different. It is a temporary helper created inside one session for one job.

I can ask Hermes to read ten blog posts and summarize them. It spins up ten subagents, each reads one, and the findings come back to the main session.

Then the subagents are gone.

Profiles separate roles and context. Subagents handle burst work inside a single conversation.

Once you have those specialist profiles, you need somewhere for them to live and a way to reach them.

11. Gateways and where Hermes runs

Messaging gateways put the same agent inside Slack, Discord and Telegram.

I have mine in Telegram, connected through a VPS, which then connects to my computer. It runs scheduled jobs and pulls research every day to give me ideas for posts.

You can run Hermes locally on your laptop. You can run it on a VPS so it works around the clock, even when your laptop is closed.

Or you can run it in a shared cloud setup so a team all accesses the same tools, context and API keys.

Where it runs depends on who needs to use it. If it is just you, local or a VPS is fine.

If you have a team, you probably want it in the cloud so everyone shares the same setup.

Wherever it runs, automation, orchestration and security decide how much of the work happens without you, and how safely.

12. Automation, orchestration and security

Scheduled tasks let Hermes begin work at a specific time without you sending a message.

I have one that runs every day at 9:02 AM. It grabs my X post from the previous day, searches YouTube for relevant podcasts and interviews on the same topic, pulls clips, and sends them to me on Telegram.

I did not write the prompt for this. I went through the workflow in a session, then told Hermes to schedule it.

Webhooks work the other way. An outside tool pings Hermes to trigger a task.

Someone hits a thank-you page, a webhook fires, and Hermes sends a notification to Slack or WhatsApp.

Subagents fan out inside one session. Kanban is different.

It works like a Trello board where tasks move through stages, and each stage has a different agent assigned. My SEO pipeline has a writer, an editor and an orchestrator.

They pass tasks between each other in the background. I only touch the approval steps.

Security stacks four layers. User authorization controls who can send requests. Tool permissions control which actions the agent can request.

Approvals pause before a risky action and wait for you to allow or reject. Hooks apply rules automatically and can block without asking.

The safest setup is to give each profile specific permissions and specific tools, and keep everything else out of reach.

How the pieces fit together

All 15 concepts are one system. The model sits at the bottom, the harness wraps around it, and the interfaces, files, tools, memory, skills, hooks, profiles, gateways and security layers sit on top of that.

Each piece stays small on its own. Together they turn one model into a working setup that does actual work for your business.

Once you know what Hermes can see, what it can remember and what you are allowing it to do, you can hand it real jobs without wondering what it will touch.

The source video covers every concept in this article on screen, with the diagrams and live demos for each one. You can watch it here: https://youtu.be/lGtBPrSrnjY

If you want to find the single highest-ROI automation opportunity in your business and have me build the Hermes Agent for you, book your free 30-minute AI audit call.

https://theaiarchitects.com/x/every-hermes-agent