Talk about AI agents usually circles around the model: which one is smartest, which one is cheapest. On its own, though, a model only answers. What turns answers into work is the program around it, the harness.
Hermes Agent, released as open source by Nous Research in February 2026, is a great way to understand this, because you can read all of it. We went through the code and the docs, and here is how it works in plain operations language, plus what we take from it.
Where this comes from
This is not an ad or a tool endorsement. Hermes is simply an open window into a kind of software that sits inside nearly every serious agent, open or closed.
What a harness is, in one sentence
A harness is the program that builds the context, offers tools to the model, runs what it asks for, records what happened and decides when to stop. The model picks the next step; the harness decides what it picks from and what it is allowed to touch.
Hermes is a harness with four clear choices:
- One simple loop. The model asks for tools, the program runs them, returns the results and repeats until a text answer comes out. Up to 500 turns per task.
- Always on. It runs on a server and answers through the terminal, Telegram, WhatsApp, Slack, email and twenty-plus other channels, with the same agent behind all of them.
- Any model. It talks to 18+ providers and switches models with one command. If one fails, it tries the next on the list.
- State in files. Conversations in a searchable SQLite database, memory in two text files and procedures in folders. You can open and read all of it.
Tools, and where they run
Hermes ships with 70+ tools: terminal, files, browser, web search, scheduling, subagents and a connection to any MCP server. Every command runs in an environment you choose: your own machine, a Docker container, a server over SSH or cloud machines that hibernate when idle.
One clever detail: the agent can write a short script that calls several tools at once. A ten-step flow becomes a single turn, which saves time and context.
Before a dangerous command, Hermes stops and asks for approval, either from a person or from a second model that rates the risk. For scheduled jobs, the default is to deny.
Small memory, growing procedures
This is where Hermes stands apart. Memory is short on purpose: an agent notes file capped at 2,200 characters and a user profile capped at 1,375. Both load at the start of each conversation and do not change mid-way, so the model answers faster and cheaper.
What grows are the skills: procedures the agent writes itself when it solves something new, makes a mistake and finds the way, or gets corrected by you. Next time, it loads the procedure only when it needs it.
- Works on the task, with tools, errors and fixes.
- Gets a nudge every ten turns to save what it learned.
- Reviews in the background, in a copy of itself, without interrupting the conversation.
- Saves the memory and the new skill.
- Goes through the weekly curator, which archives anything unused for 30 days.
How Nous makes money from free code
The code is free (MIT). Revenue comes from an optional account, Nous Portal: one subscription that bundles hundreds of models, tools like search and image generation, and ready-made servers for the agent, instead of juggling five different API keys.
According to the press, the company reached about $36 million in annualized revenue in September and raised $90 million in October 2026. It is the “give away the agent, charge for convenience” model, and it shows that value is shifting toward making agents work, and away from the model itself.
What to watch before putting an agent like this into operations
- Isolate it. An agent with a terminal runs in a container, never loose on someone’s machine.
- Approve what matters. Writes to external systems, payments or messages to customers go through a person.
- Check the memory. The docs themselves warn that weak models may say “saved” without saving. Open the file.
- One agent per workspace. Two agents in the same folder mix memories nobody wrote.
- Mind the data. A model account in another country needs a contract and a legal basis for the transfer (GDPR and similar laws).
What this changes in practice
The Hermes lesson is that a good agent is a model plus a well-designed harness: a simple loop, tools with limits, short memory, procedures that improve with use and rules enforced at execution time. The same design works for an agent that reconciles accounts, drafts proposals or answers customers.
- The AI Analyst maps the recurring tasks and writes the procedure for each one, which becomes the agent’s skill.
- The Forward Deployed Engineer builds the harness inside your operation: an isolated environment, connectors, approvals and memory.
- Training teaches the team to read what the agent learned and correct it when needed.
Sales, finance, support and purchasing are just examples. If a task repeats, it can be automated, and the harness is what makes that automation learn over time.
The model answers. The harness is what turns answers into finished work, and that is where the difference lives.
Enjoyed this article?
Get the next one before it hits the blog.
A short email per issue with what actually matters about AI agents at work. No spam.