Hermes on the laptop: what an agent profile is

What I wanted to fix

In the previous post, I shared how I built a team of specialized agents inside Grok Bot to split up daily tasks.

Having focused assistants on my phone helped a lot, but using them solely through a chat app quickly hit practical limits. Conversation history was trapped inside the chat screen, connecting those conversations to files on my computer was awkward, and I had zero control over how agents stored memories or executed local tools.

I work during the day, and at night and on Sundays I build with AI agents. My free time is limited, so every tool I use needs to solve real problems. In mid-September 2026, I decided to take the next step: move agents out of a locked chat tab and run them directly on my laptop.

I wanted to test agents that could read local directories, run tools, and keep independent identities without mixing up topics. That is when I installed Hermes Agent.

Quick summary

• In the previous post, I shared how I built a team of specialized agents inside Grok Bot to split up daily tasks.
• In mid-September 2026, I installed Hermes Agent on my laptop to run agents directly on my computer.
• Each agent operates from its own profile, with isolated instructions, tools, and stored memories.
• The so-called soul of an agent is simply a plain text file containing practical rules, clear boundaries, and a defined role.
• On my laptop, I set up Optimus as the leader of local agents and a local Security agent to protect the machine itself.

First: what Hermes Agent and an agent profile are

If you have only used AI in a browser tab or a phone app, running an agent directly on your computer feels very different at first. Two concepts explain what changes.

Hermes Agent is an open-source software program designed to manage and run artificial intelligence agents. Instead of depending on a commercial chat window, you run the software right on your own machine. It lets the underlying model use command-line tools, read and write files, and keep persistent records of its work. When it runs locally on a personal computer, we call it Hermes Desktop.

An agent profile is the isolated environment where each specific assistant lives inside that program.

Think of user accounts on your laptop or separate viewer profiles on a family streaming service. When you create one profile for yourself and another for a guest, your search history, watch lists, and preferences never show up on their screen. Each person opens their own customized workspace.

With AI agents, the logic is similar. Each profile maintains its own directory on disk, its own conversation logs, its own memory notes, and its own master instruction file. When you launch a profile, the model loads only that specific bundle of information. It never sees instructions belonging to other agents.

The comparison stops there: a person can leave their desk and still remember what their family talked about over lunch. An artificial intelligence model lacks that continuous personal consciousness. It only knows what is explicitly loaded into its context window for that active session. If a detail is not written down in that profile’s instruction file or memory records, to the model that detail simply does not exist.

This instruction file is often called the agent’s soul. The word sounds poetic, but the purpose is strictly practical. The soul is a plain text file where you state who the agent is, what tasks it is permitted to execute, what tools it may touch, and what actions it is strictly forbidden from taking.

What I used

During this stage in mid-September 2026, I used:

• A standard personal laptop.
• Hermes Agent (Hermes Desktop), installed locally on the machine.
• AI model access connected through API credentials to handle reasoning and instructions.

I did not purchase new servers or build complex cloud setups at this point. All configuration and testing happened right on the laptop I open in the evenings.

How I structured the profiles on the laptop

To avoid repeating the mistake of piling conflicting chores into one window, I built my local setup around distinct profiles. Inside Hermes on my laptop, I established two primary agents:

1. Optimus: the leader of local agents on the laptop. Optimus receives my requests on the computer, organizes workflows that need to run locally, and coordinates tasks on the machine.
2. Local Security: watches over the security of the laptop itself. It checks file permissions, reviews scripts before sensitive operations run, and ensures that dangerous system changes never happen without my explicit approval.

Configuring each profile followed a clean layout inside its instruction file:

• Core role: a single sentence defining what the agent exists to do.
• Action rules: a short checklist of steps the agent must take when receiving a request.
• Boundaries and limits: a clear list of what the agent cannot touch, delete, or decide on its own.
• Output format: instructions to answer in brief bullet points, without introductory filler or small talk.

With both profiles operating independently, Optimus focused on task coordination while local Security guarded the integrity of my computer.

What went wrong and how I fixed it

My first mistake was trying to write long, overly elaborate prompts for each agent’s soul.

I thought I needed to paint deep personality traits and philosophical explanations so the assistant would take its job seriously. The real outcome was the exact opposite of what I wanted: the model consumed precious context tokens on decorative adjectives, produced dramatic introductory speeches before answering, and took forever to get to the point.

I quickly learned that an agent’s instruction file does not need literary flair. It needs operational clarity. I replaced those wordy paragraphs with direct bullet points: who the agent is, what it handles, what it never touches, and how its output should look.

If you tell an agent to “be a helpful and inspiring partner”, you give the machine zero useful guidance. When you tell an agent to “reply in short bullet points, use active verbs, and never delete files without confirmation”, you give it an exact standard to follow.

The lesson was direct: an agent’s soul is a practical work manual, not creative writing. Start with a few tight rules, and add new ones only when you notice a specific, recurring mistake during real work.

Costs and limits

Because Hermes Agent is an open-source project, installing it on my computer cost nothing in software licenses. The only monetary cost was the token usage from the underlying language models used to process prompts and answers during my testing sessions.

However, the practical limits of running agents locally on a laptop showed up almost immediately:

• The agents only run while the laptop is powered on, awake, and connected to the internet. If the lid closes, or if the machine sleeps or drops its Wi-Fi connection, the agents stop working.
• I could not schedule autonomous tasks to run overnight or throughout the day while I was away at my day job.
• Running local tools and multi-step reasoning draws battery and system memory, which meant the laptop needed to stay plugged into a wall outlet for longer sessions.

What I would do differently

If I were setting this up again from scratch, I would write far shorter instruction files right from the start. I would not waste evenings trying to give an agent subtle psychological nuances when all it needed was three or four strict operational rules.

I would also standardize how I test new profiles from day one. Creating a dedicated test profile for experiments prevents temporary trial settings from cluttering up your primary agent’s instructions.

Who this is for and who it is not for

Running Hermes on your laptop with dedicated agent profiles is a good fit for:

• Anyone who wants direct control over where their agent files, logs, and stored memories reside.
• People who want strict separation between distinct workflows, ensuring one agent never confuses its instructions with another.
• Anyone willing to learn how to write clean instruction files in plain text to see how specific rules change an agent’s output.

It is not a good fit for:

• Anyone looking for 24/7 background automations without keeping a computer running at home.
• People who prefer one-click mobile apps and do not want to manage local files or terminal settings.
• Anyone expecting an AI assistant to make autonomous decisions without human review.

Wrapping up

Setting up Hermes Agent on my laptop and creating independent profiles fundamentally changed how I work with artificial intelligence. Having Optimus coordinate local tasks while local Security checked every critical action gave real structure to my evening study sessions.

Yet the limitation of needing the physical laptop running all the time quickly forced me to look for another solution for tasks that had to run while I was away.

In the next part, I will show the next step in this evolution: Hermes Cloud: agentes que trabalham com o notebook desligado.

If you are also experimenting with setting up local agents and want to share practical notes, join me on X: @RodrigoesPorto.

Sources and links

• Official Hermes Agent documentation: https://hermes-agent.nousresearch.com/docs
• My official profile on X: https://x.com/RodrigoesPorto