Autonomous agents: an advanced option and its risks

In testingChecked on 24/10/2026 (French version)Translated on 27/10/2026

This lesson is still being tested: it improves with learners' feedback. If something is unclear, it is the lesson that needs fixing, not you.

Follow this lesson with Claude Code

With the course companion installed, open Claude Code in your working folder and type:

/formation-ia:suivre m09-l06 en

It reads the up-to-date lesson on this site and guides you step by step, without changing anything without your consent. Install the companion. Claude Code is not included in the free plan: an eligible Claude subscription, or a Console account billed by usage for the terminal.

By the end of this lesson, you will be able to

say what an autonomous agent is, and what sets it apart from an agent like Claude Code in Manual mode;
know the risks specific to these agents, through a real example: Hermes Agent;
read the security settings to look at before any installation;
decide, on paper, whether you really need one.

Why it matters to you

An autonomous agent promises a lot: it answers on your messaging apps, keeps the memory of your projects, runs tasks at night. But it also acts when you are not watching, with the access you gave it, and sometimes without asking for your approval. This lesson does not make you install anything. It teaches you to read what such a tool can do, and to decide with full knowledge. For most needs, the scheduled tasks from module 8 are enough.

Before you start


You need	Nothing to install: the lesson is done on paper. Your task sheet and your stop plan from module 8
Cost	€0 for the lesson. Hermes Agent is free software (MIT license); the AI model it uses may be paid
Duration	35 minutes
Reminder	An autonomous agent acts with the access you give it, when you are not watching

Understand it in 3 minutes

What changes with an autonomous agent:

	Claude Code in Manual mode	An autonomous agent
When it acts	While you work with it, or at the time of a scheduled task that you set up (module 8)	Also when you are not there: scheduled tasks, messages received
Who can talk to it	You, on your computer (unless you connected a messaging app such as Telegram or Discord)	Depending on its settings: you, or other people, through a messaging app
Your approval	Asked before most actions that edit a file, run a command or touch the network	Depending on its settings, sometimes only for certain commands

A real example: Hermes Agent, from Nous Research, according to its website and its documentation (24/10/2026):

it connects to messaging apps (Telegram, Discord, Slack, WhatsApp, Signal, email), keeps a memory of your projects, and runs scheduled tasks unattended;
the software is free (MIT license), but it needs an AI model: for example a Nous Research subscription, your ChatGPT subscription (Codex models), Claude (a Max plan with usage credits, or an API key billed by usage), or OpenRouter (an API key);
there is an app for Mac and Windows.

What its security page says:

Setting	What it does	To start with
Approval requests (`approvals.mode`)	By default “smart”: an AI judges the risk of commands flagged as dangerous; “manual” always asks for your approval for those commands; “off” stops asking anything	`manual`. And know that only commands flagged as dangerous ask for your approval
YOLO mode	Removes all approval requests for the session, except for a short list of commands that are always blocked. It is saved with the session: it is still on if you reopen it	Never
Where commands run (`terminal.backend`)	“local”: directly on your computer, with no isolation; Docker, Modal or Daytona: in an isolated space. Warning: in an isolated space, Hermes skips its usual check for dangerous commands, because the isolation is what protects you; the `manual` setting changes nothing there, only your own blocking rules still apply	An isolated space, or a separate computer, checking what you share with it (folders, keys)
Who can write to it	Without an allowlist, the messaging app refuses everyone; `GATEWAY_ALLOW_ALL_USERS=true` opens it to everyone	Only your own identity in the list. Never “everyone”
Scheduled tasks (`cron_mode`)	By default, a dangerous command in a scheduled task is refused	Keep this setting
Keys	Stored in `~/.hermes/.env`; the documentation recommends restricting read permissions on it (on Mac and Linux: `chmod 600 ~/.hermes/.env`)	Never in a conversation
Existing sign-ins (`auth.adopt_external_logins`)	By default, when it has no sign-in of its own, Hermes can take over those of Codex and Claude Code installed on the same computer, and renew them	Give it its own sign-in, and set this setting to `false`

Do it together

1Step 1: Read before installing

Open the official security page: hermes-agent.nousresearch.com/docs/user-guide/security. For each of the seven settings in the table, find the passage that talks about it. You do not need to understand everything: you are looking for what acts without you.

2Step 2: The “autonomous agent” sheet, on paper

Copy it and fill it in:

AUTONOMOUS AGENT SHEET
The need: [what I would like it to do]
Would a scheduled task from module 8 be enough? [yes / no, because…]
What it could read: [folders, accounts]
What it could do: [write, send, delete, pay]
Who could write to it: [only me / others]
Where its commands would run: [my computer / an isolated space / a separate computer]
What the AI model would cost, per month: [amount, and where I cap it]
How I stop it: [the exact procedure, found in the official documentation]
My decision: [install later / do not install], on [date]

If the “How I stop it” line stays empty, the decision is already made: do not install.

3Step 3: Compare with what you already have

Go back to your routine from module 8 (point-semaine-auto). It reads your journal, in Manual mode, on your computer, and you know how to stop it in two minutes. For most reporting and reminder needs, that is enough.

Your turn

Fill in the “autonomous agent” sheet for a real need, or an imagined one.
If you decide to install one day: follow only the official documentation, with a dedicated account and folder, approvals.mode set to manual, never YOLO, only your own identity allowed, and the stop plan filled in before the first launch.
File the sheet in your journal.

Check your work

It worked if:

you can say what the seven settings in the table do;
your sheet is filled in, with a dated decision;
the “How I stop it” line is filled in, or your decision is “do not install”;
you installed nothing during this lesson.

Common mistakes:

installing from a video, without reading the security page;
leaving the “local” setting on the computer where all your documents are;
opening the agent’s messaging app to everyone;
turning on YOLO “so that it stops asking”;
forgetting the cost of the AI model, which also runs at night.

Keep it and pick it up later

Keep the “autonomous agent” sheet in your journal, next to your task sheets.
If you ever install an autonomous agent, add it to your stop plan (module 8, lesson 6) the same day.

Going further

optional

Hermes’s documentation recommends, for permanent use, an isolated space or a separate computer, an allowlist, never running it as administrator, and regular updates (hermes update).

New words

4 words explained

Word	What it means
Autonomous agent	An agent that also acts without you: scheduled tasks, messages received
Isolated space (container)	An environment separate from your computer, where commands do not touch your files, except the ones you share into it
Allowlist	The only people who can write to the agent
YOLO	A mode that removes approval requests. Never to be used to start with

Check your understanding

3 questions, no grade. An explanation appears after each answer.

1In Hermes Agent, what does "YOLO" mode do?

It runs commands in an isolated space, so that dangerous commands do not touch your computerIt removes the approval requests before dangerous commands, for the whole sessionIt asks for your approval before every command, even those not flagged as dangerous

2With the "local" setting, where do Hermes's commands run?

Directly on your computer, with no isolationIn an isolated space, separate from your filesOn Nous Research's servers, remotely

3You cannot find, in the documentation, how to stop the agent you want to install. What do you do?

I do not install it until I know how to stop itI install it, and I ask the agent itself how to stop itI install it in manual mode: it will ask before acting

What's next?

Lesson 7 · Safeguards: loops, costs, data: the three risks of any work with several AIs, and how to keep them in check.

Module 9:

✏️ Do the exercises

My space: your progress, your exercises and your last lesson.

To be checked: Hermes Agent has been neither installed nor tested for this lesson; the settings are described from the documentation, and may change from one version to another. For stopping it, the commands reference page (hermes-agent.nousresearch.com/docs/reference/cli-commands) gives hermes gateway stop for the messaging app, hermes cron pause or hermes cron remove for a scheduled task, and hermes uninstall to remove the software; these commands have not been tried, so check their names in the documentation for your version.

Sources and checks

4 sources

Date, version, sources Written on 24/10/2026. Checks of 24/10/2026:

presentation of Hermes Agent (messaging apps, memory, unattended tasks, Mac and Windows app): hermes-agent.nousresearch.com;
MIT license: LICENSE file of the NousResearch/hermes-agent repository;
possible AI models (Nous Portal, ChatGPT or Codex, Claude with a Max plan and usage credits or an API key, OpenRouter): hermes-agent.nousresearch.com/docs/getting-started/quickstart;
approval requests, YOLO, “local” setting, isolated spaces, allowlist, scheduled tasks, keys, advice for permanent use: hermes-agent.nousresearch.com/docs/user-guide/security;
stop commands (hermes gateway stop, hermes cron pause, hermes cron remove, hermes uninstall): hermes-agent.nousresearch.com/docs/reference/cli-commands.

Sources checked: