ع
🤖 AI Agents

Hermes Agent for Beginners: First Commands and How It Thinks

You've heard what agents can do — now run one. The agent loop, the first five commands worth learning, and three starter tasks that show what 'it did it for me' actually feels like.

Beginner9 min read

What you'll be able to do by the end

  • Run your first Hermes conversation and know what each part of the screen means
  • Explain the plan → tool → check loop the agent runs on every task
  • Use the handful of commands and slash commands that cover daily use
  • Complete three real starter tasks end to end

Before you start

  • Hermes installed and chatting — if not, start with our Desktop setup guide

Reading about agents is like reading about bicycles. The idea only lands when you ride. This article gets you on the bike: your first session, the mental model for how the agent works, the small set of commands that matter, and three tasks you can complete today.

First contact

Install done? Open a terminal and type hermes. (Nothing installed yet? Our Hermes Desktop setup guide walks you through it in minutes.)

You’ll see a welcome banner: your model, your tools, your skills. Now write something concrete:

Summarize this repo in 5 bullets and tell me what the main entrypoint is.

Watch what happens. Before answering, Hermes may list files, read a few, maybe run a quick command — each call visible as it happens — then compose the answer. That visible chain of tool calls is the agent working. Nothing happens behind a curtain.

The loop: how an agent thinks

Every agent task runs the same engine:

  1. Understand the goal.
  2. Plan the steps.
  3. Act by calling a tool — terminal, file operations, web search, browser.
  4. Check the result honestly.
  5. Adjust when reality disagrees with the plan.
  6. Deliver when done.

Step 5 is the magic. A script written yesterday does exactly what it said and fails exactly where it said. An agent reads the error message, forms a new hypothesis, and tries again — the way you would, minus the coffee breaks. And because every tool call prints to the transcript, you supervise by scrolling, not by hovering.

Your job as the human is different from prompting a chatbot: state goals clearly, review plans before approving changes, and correct course in plain language when needed. Less typing, more managing.

The commands worth knowing

Five minutes of vocabulary covers ninety percent of daily use:

Command What it does
hermes Start an interactive chat session
hermes chat -q "..." One-shot question, no session opens
hermes --continue / -c Resume your most recent session
hermes model Switch provider or model at any time
hermes doctor Diagnose problems and get exact fixes

Inside any session, type / for autocomplete: /help, /tools, /model, /save are the ones you’ll actually use. Multi-line input: press Alt+Enter or Ctrl+J for a new line. To interrupt mid-task, just start typing — your message redirects the agent; Ctrl+C also works.

Three starter tasks, end to end

1. File cleanup (safe version). “List my Downloads folder grouped by type and propose a cleanup plan — delete nothing.” You see how it reasons while keeping your files untouched. Approve the plan later once you trust it.

2. Document summaries. “Read every PDF in ~/documents/reports and give me 3 bullets per file plus combined action items.” The agent loops through the folder itself — the multi-file workflow that used to mean an afternoon of copy-paste.

3. Research brief with web sources. “Compare the top 3 project management tools for a 5-person team from current sources, save to comparison.md.” Search, reading, judgment, and a written artifact — then check the file yourself. Verification stays your habit even when the agent is fast.

Run those three and you’ve covered the pattern behind most real work: inspect something, transform something, produce something.

From here

Next steps in the series: put it on your phone with Hermes + Telegram, understand installation deeply in the Desktop setup guide, or jump ahead to skills, model choice, and scheduling in Skills, Models & Automation.

For background on why agents differ from the chatbots you may already use, start with What Is Hermes Agent? — and for chat-side technique, How to Use Claude pairs well.

Steps

  1. Step 1: Start your first session

    Open a terminal (or the desktop app) and type `hermes`. A welcome banner appears showing your current model, available tools, and skills. Type a normal sentence and press Enter — that's already an agent request.

    Note: Prefer a richer interface? `hermes --tui` opens the modern TUI with menus and mouse support. Both share the same sessions and settings.

  2. Step 2: Give it a goal, not a step list

    Write outcomes: 'Summarize this repo in 5 bullets and tell me the main entrypoint' or 'Check my disk usage and show the 5 largest folders'. Notice what happens next: Hermes replies with a short **plan**, then calls tools — terminal, file reads, search — each shown in the transcript as it runs.

    Note: If a task is taking long or going sideways, just type your new instruction — it interrupts the current work and switches to it. Ctrl+C also stops.

  3. Step 3: Watch the loop: plan → tool → check

    Every task follows the same cycle: read the request, plan steps, call a tool, inspect the result, adjust, repeat until done. When a command fails, Hermes reads the error and corrects itself — this self-correction is why agents feel different from chatbots. You'll see every tool call in the transcript, so nothing happens invisibly.

    Note: When output disappoints, don't restart — reply with a specific correction like a manager would: 'only PDFs from March', 'keep the original names'.

  4. Step 4: Learn the essential commands

    A short list covers daily life: `hermes` starts interactive chat; `hermes chat -q "..."` fires one question without opening a session; `hermes --continue` (short `-c`) resumes your last session; `/help`, `/tools`, `/model`, `/save` are the slash commands worth memorizing inside a session; `hermes doctor` diagnoses any problem; `hermes model` changes provider/model anytime.

    Note: `hermes --continue` is the one beginners miss: sessions persist, so yesterday's context comes back exactly where you left it.

  5. Step 5: Starter task 1 — file cleanup

    Try: "List everything in my Downloads folder, group by type, and delete nothing — give me a proposed cleanup plan." Reviewing the *plan* before allowing changes is the right beginner habit: you see how it reasons, with zero risk.

    Note: Once comfortable, allow the actual move: 'sort into subfolders by type'. Same skill, now with effects.

  6. Step 6: Starter task 2 — summarize documents

    Drop files into a folder, then: "Read every PDF in ~/documents/reports and give me a 3-bullet summary of each plus a combined list of action items." This shows multi-file reasoning: the agent loops through files itself, no manual uploading.

    Note: Scanned (image-only) PDFs read poorly — ask it to tell you which files were scans so you can handle them separately.

  7. Step 7: Starter task 3 — research brief

    "Compare the top 3 project management tools for a 5-person team, using current web sources, and save the comparison to comparison.md." Watch it search, read pages, weigh trade-offs, and write the file — then open the file to verify. Verifying outputs is part of working with agents, not distrust.

    Note: One-shot mode (`hermes chat -q "..."`) is perfect here: fire the task and come back to the finished file.

Common mistakes — and how to avoid them

MistakeTyping step-by-step instructions ('first do X, then Y') like scripting a robot.

Do this insteadState the outcome and constraints. The agent plans better than your script when it knows what 'done' means.

MistakeRestarting the session when a result misses the mark.

Do this insteadCorrect within the same session — accumulated context is what makes the second attempt sharper than the first.

MistakeApproving every scary-looking command blindly, or denying everything out of fear.

Do this insteadRead what the command does. Dangerous actions require your explicit approval — that prompt is a feature, not an obstacle.

MistakeAssuming a number, date, or source in the output is correct.

Do this insteadVerify facts that matter against the original source before you act on them. Agents are fast, not infallible.

❓ Frequently asked questions

Do I need to memorize commands?

No — natural language covers almost everything. The commands above are conveniences: starting, resuming, one-shot queries, and switching models. Inside a session, type / to see the full autocomplete menu.

What happens if the agent gets stuck?

It tells you. Failed tools and errors appear in the transcript along with its recovery attempts. If it truly can't proceed, it says what's blocking it — missing access, unclear instructions, unavailable tool — and you adjust. You're never left staring at silence.

Can it break my computer?

Potentially destructive actions trigger an approval prompt first, so nothing irreversible happens silently. Keep that habit: review plans before allowing changes, especially on folders you care about, and consider backups for anything important.

Does it remember previous conversations?

Sessions persist — resume the last one with hermes --continue or a specific saved session. Beyond sessions, Hermes keeps persistent memory about your preferences and environment across conversations, which is why it feels increasingly 'yours' over time.

Can I use it from my phone?

Yes — connect Telegram (or another messaging platform) and the full agent answers from your pocket. See our Telegram connection guide in this series.