Build Your Own Jarvis

From chatbot to agent

A chatbot answers a question. An agent remembers, uses tools, and gets things done. Students build one with Hermes Agent, a free, open-source agent from Nous Research. It's what my own Jarvis runs on.

💛

A soul A personality file the student writes: who it is, how it talks, and its rules.

🧠

A memory It remembers what the student chooses to teach it, across conversations.

🛠️

Skills Step-by-step abilities it learns, like "turn my notes into flashcards."

⏰

A schedule Tasks it does by itself, like a morning study briefing before a test.

Why it works as a first workshop: every age gets the same "wow" moment of an AI that answers to its name. From there, each part (soul, memory, skills, and schedule) becomes a deeper project students can grow over a term.

How I built mine

My Jarvis runs on a Mac mini at home, always on, and wakes up when I say "Hey Jarvis." Students build a scaled-down version of the same design.

💛

A personality file One file sets its name, its tone, how short its answers are, and what it never does without asking.

🎙️

"Hey Jarvis" A small local listener (openWakeWord) hears the name and hands off to Hermes. The mic stays local until then.

🔌

Apps, carefully Calendar, notes, and email connected with only the tools it needs. Email starts read-only: it briefs me before it ever sends.

📱

Always reachable It runs as a background service, so the same assistant, with the same memory, answers from my phone.

📈

It learns my routines Repeat a workflow, like a morning brief, and it writes a reusable skill. I review what it learned.

The lesson for students: the most important part of my setup isn't a feature. It's the permissions: what it may read, what it must ask before doing, and what it never touches. That's the core of this workshop. I wrote about the setup on LinkedIn.

What the first session looks like

60 minutes, built around a live "Hey Jarvis" moment. It can be shortened to 30 minutes by skipping the build step for younger groups.

0–5 · "Hey Jarvis": I wake a classroom copy of my Jarvis on the projector, and it answers the students' questions live. The classroom copy has no personal accounts connected.
5–15 · How it works: the four parts of every agent (soul, memory, skills, and schedule), taught through a game. One student plays "Jarvis" and must follow a rules card exactly, while the class tries to make it break a rule.
15–35 · Design your Jarvis: each student chooses a name, a personality, three jobs, three rules, and a wake word. Ages 8–12 use a paper design card, and older students write a real personality file.
35–50 · Bring it to life: I load a few students' designs into the classroom agent live, so they hear their own Jarvis answer. Teenagers test theirs in a safe sandbox.
50–60 · Show and what's next: students show their designs and add the project to their portfolio. Next sessions add memory, skills, and a schedule.

The concept

The same four questions as the other concepts.

Student outcome

Each student leaves with their own working AI agent and a clear picture of how it works inside. It has a name and personality they wrote, a memory they control, at least one skill they taught it, and one task it runs on its own schedule. For example, a Jarvis that quizzes them every evening before an exam, or one that plans their week around practices and assignments.

Why families would love it

"AI agents" are the next wave after chatbots. Students who understand how they work from the inside, including memory, tools, and permissions, will use them more wisely and build with them sooner.
It's genuinely useful. A study agent built around the student's own schedule and notes helps with schoolwork without doing the work for them.
It's built on open-source software that professionals use, not a toy. That makes it a serious item for a portfolio, a résumé, or a university application.

How it grows into a deeper project

Give it a soul: write the personality file, including its name, voice, and the rules it never breaks.
Decide what it remembers: choose what it should keep, what it must never store, and how to make it forget.
Teach it a skill: write a skill file for one task, test it, and improve it when it gets something wrong.
Put it on a schedule: automate one helpful task, and decide what it may do without asking first.
Give it a voice (stretch goal): add a "Hey Jarvis" wake word with a small listener that runs on the computer itself, so nothing is sent anywhere until it hears its name.
Demo and share: present Jarvis to a mentor, and publish the best skill as an open-source contribution others can use.

Students' own ideas and decisions

Students decide what their Jarvis is for, who it is, and what it's allowed to do. The most important decisions are about trust: which tools it can use, what it remembers about them, and what it must ask before doing. These are the same questions AI companies face. Students answer them for their own agent and explain their reasoning in their portfolio write-up.

What a student writes: a sample soul

The personality file is plain language. It's the student's design, written in their own words.

# Who you are You are Atlas, my study partner for Grade 11. You are calm, a little funny, and you never lecture. # What you help with - Planning my week around basketball and assignments - Quizzing me before tests, using my class notes # Rules you never break - Give hints, not answers, on homework - Never store my full name, school, or address - Ask me before you add anything to my schedule - If you're not sure, say so # Written by the student · checked by the instructor

How it adapts by age

Every age gets the "Hey Jarvis" moment and designs their own agent. The hands-on build is for teenagers, and younger students take part without any accounts or installs.

Age	What students do
8–12	Watch a live demo of the instructor's Jarvis, then design their own agent's soul on a paper card: name, personality, three jobs, three rules. The class votes on which rules matter most.
13–15	Guided build in a shared classroom setup. Students write the soul and one skill, and test them with the instructor's agent.
16–18	Full build: their own agent in a safe sandbox, with memory, a custom skill, a scheduled task, and a written explanation of their trust decisions.

Safety setup

Agents can take real actions, so the setup matters more here than in any other workshop.

Sandboxed: student agents run in an isolated environment (Hermes supports Docker and cloud sandboxes), never on a family computer with access to personal files.
No personal accounts connected: no email, social media, or payment accounts. The agent only works with what the student gives it in class.
Instructor-managed AI access: students never handle API keys. Usage is capped and monitored.
Permissions are a lesson: students decide and document what their agent may do without asking. The default is "ask first."
Voice stays local: if students add a wake word, the listener runs on the device and the microphone only connects to the agent after the wake word.
Consent: the hands-on build is for ages 13+ with parental consent, and students use first names only.