How to learn creating AI agents has a short verdict, which is that you should learn to design one agent before you learn any tool.

I'm Julian Goldie, and I test agent tools on camera most days for 400,000+ YouTube subscribers.

After all that testing, my view is that the tool matters far less than the nine choices you make inside it.

Here's the path, with my pick at each step.

  1. You learn the nine design choices, and my verdict on each is in the table below.
  2. You fill in a one-page agent brief for one small job.
  3. You create the agent in StarNet if you want to see every choice, or in Hermes Agent if you want it to do real work.
  4. You test it on five real tasks.
  5. You break it, fix it and create the next one.

This page is a reference, so you can jump to the choice you're stuck on.

The Verdict On How To Learn Creating AI Agents

The fastest way to learn is to create one small agent with safe settings and then try to break it.

These are the settings I'd choose, rated by how much each one matters for a beginner.

Design choice My beginner pick How much it matters Verdict in one line
The job One clear outcome 10/10 A vague job ruins everything after it.
The instructions Five to ten plain sentences 9/10 Clear rules beat a clever model.
Approvals Ask for approval 9/10 This is your safety net while you learn.
Permissions Read only 8/10 An agent can't damage what it can't change.
The test Five real tasks 8/10 You can't improve what you haven't scored.
Tools One or two 7/10 Fewer tools mean fewer wrong turns.
The brain Free or already paid for 6/10 A beginner job rarely needs the top model.
The trigger Chat 6/10 You should watch it before you schedule it.
Memory Built in 4/10 The default is enough for a first agent.
Persona Composed 2/10 Tone is fun to set, and it's the least important choice.

The ratings are my own opinion from creating agents daily, and they aren't a benchmark.

The order surprises people, because the brain sits in the bottom half.

Beginners spend most of their time choosing a model and almost none writing the job.

I'd flip that.

Where To Create Your First Agent: StarNet Vs Hermes Agent

I've created agents in both of these, and both are free and open source.

What I'm rating StarNet Hermes Agent
How you create an agent You click through a creation screen. You run three terminal commands.
How visible the design choices are 9/10 6/10
How practical it is for real work 5/10 9/10
Triggers I've used Chat Chat, schedules and messages
Best for Learning what an agent is made of Running an agent every day
My verdict It's the best classroom. It's the best workshop.

I scored StarNet low on practical work because of what I said in my own video.

I think it's fun to play with, and I don't think there's a massive amount of practical use cases yet.

I scored it high for learning because the name, personality, approval level, brain and reasoning level are all buttons on one screen.

Hermes is the opposite, because the choices live in commands and files, and the payoff is an agent that can run on a schedule.

My StarNet AI agent review goes through the station feature by feature.

🔥 Want my exact How To Learn Creating AI Agents setup, step by step? Inside the AI Profit Boardroom you get step-by-step video trainings, my installable Agent OS and four coaching calls a week with 3,400+ members building real automations. → Join the AI Profit Boardroom

Design Choice 1: The Job

The job is the single outcome the agent delivers.

Option Example Verdict
A broad job "Help with my marketing." You should avoid it, because nobody can check it.
A narrow job "Name three keywords from the queries I paste in." You should pick this.
Several jobs in one agent "Do research, writing and email." You should split it into several agents.

I teach one agent doing one job.

If the job doesn't fit in one sentence, the agent isn't ready.

Design Choice 2: The Brain

The brain is the model that reads the instructions and decides what happens next.

Option What it costs you Verdict
A free model Nothing This is the best place to learn.
A subscription you already own No new bill This is the best everyday choice.
A pay-per-use API key You pay for what the agent uses This is best once you know what the job needs.

In StarNet I signed in with my Grok account in one quick step.

I then picked the model and the reasoning level, which runs from low through medium and high up to max.

I chose Grok 4.6 over Grok 4.7, because I actually prefer Grok 4.6 for that kind of chat.

In Hermes I keep one profile per model, so I can compare brains on the same job.

My list of the best Hermes Agent model picks covers the brains I use.

Design Choice 3: The Instructions And Persona

The instructions are the rules, and the persona is the tone.

Part Where it lives in StarNet Where it lives in Hermes Verdict
Persona You pick a personality on the creation screen. It sits in the profile's SOUL.md file. You should set it once and move on.
Instructions You write them in the instructions field on the detail panel. You write them in the same SOUL.md file. You should spend most of your time here.

In my StarNet video I talked through four personalities, which were warm, dry, composed and unhinged.

You can change the personality at any time.

Good instructions say who the agent is, what the job is, what finished looks like and what it must never do.

They also say what to do when the agent is unsure, which should be to stop and ask you.

Design Choice 4: Tools And Permissions

Tools are what the agent can use, and permissions are what it's allowed to touch.

Setting Beginner pick Verdict
Number of tools One or two Every extra tool is one more way to go wrong.
Web access You allow it when the job needs research. It's low risk for a first agent.
Local file access You leave it off at first. You turn it on only when a test needs it.
Connected apps You connect none at first. You add one app after the agent passes its tests.

StarNet's app catalog is labelled "add an ability", and it includes Gmail, Notion and GitHub.

Its detail panel shows access permissions for things such as the web and local files, and you can allow or remove each one.

My rule is that an agent gets only the access its job needs.

Design Choice 5: Memory

Memory is what the agent keeps between sessions.

Option What it is Verdict
Built-in memory The tool stores a small set of notes for you. You should pick this first.
A memory add-on A plugin or a separate notes system holds more. You should leave this until you hit a real limit.

StarNet gives each agent a Memories tab that shows what it has remembered.

Hermes ships with two files, which are MEMORY.md for the agent's own notes and USER.md for your preferences.

You should read the memory once a week, because it shows you what the agent thinks matters.

Design Choice 6: The Trigger

The trigger is what starts the agent.

Trigger How it works in Hermes Verdict
Chat You run hermes -p and the profile name. You should start every agent here.
Schedule You run hermes cron create with a schedule and a prompt. You should add it after the agent passes its tests.
Message You run hermes gateway setup to connect an app such as Telegram. You should add it when you want to reach the agent away from your desk.

Chat is the best trigger for learning, because you see every reply as it happens.

Design Choice 7: Approvals And Guardrails

Approvals decide when the agent must check with you.

Option What it means Verdict
Ask for approval The agent checks with you before it acts. You should start here every time.
Full power The agent gets on with the job. You should switch to it only after you trust the agent.

Those are the two modes StarNet offers when you create an agent.

I'd watch a new agent for a few sessions before I loosened the leash.

Hermes still asks before anything risky, and its hermes pause command is an emergency stop for scheduled jobs.

Design Choices 8 And 9: Testing And Improving

Testing is five real tasks, each scored as good, fixable or wrong.

Improving is changing one thing and running the same five tasks again.

Habit Verdict
You test with made-up tasks. It's weak, because real work is messier.
You test with five tasks from your own week. This is the one to pick.
You change several settings at once. You should avoid it, because you won't know what helped.
You change one setting and re-test. This is the one to pick.

🔥 Want the agent setups behind these verdicts? Inside the AI Profit Boardroom, you get my installable Agent OS and step-by-step courses on Hermes Agent, Grok Bot, DeepSeek Harness and Claude Opus 5.5. There are 3,400+ members and four coaching calls a week. → Get access here

The Fill-In Agent Brief

You copy this template and complete it before you create anything.

AGENT BRIEF
Name:
The one job:
Finished looks like:
Brain and reasoning level:
Personality:
Instructions (5-10 sentences):
Never do:
Tools (max 2):
Permissions:
Memory:
Trigger:
Approval level:
Five test tasks:

Worked Example: One Agent, Created End To End

Here's the brief for a small SEO helper, which matches what I told my own StarNet agent I wanted.

Brief line Example answer
Name Scout, which is an example name
The one job It names the three keywords I should write about from the queries I paste in.
Brain and reasoning level I use Grok on medium reasoning.
Personality I choose composed.
Never do It never invents a number and never publishes anything.
Tools It has web search only.
Permissions It can read the web and can't touch local files.
Trigger I start it from chat.
Approval level It asks for approval.

In StarNet, you name the agent, check its appearance, pick the personality and choose the approval level.

You then connect the brain by signing in or pasting an API key, and you pick the model and the reasoning level.

The agent is created in one or two clicks, and a "waking up" animation plays.

Mine said "I'm awake. Let's give the station a direction."

It then asked me what made me want to set up an AI agent, and I told it SEO and gave it aiprofitboardroom.com.

In Hermes, the same agent takes three commands.

hermes profile create scout
hermes -p scout setup
hermes -p scout

The first command makes an isolated profile, the second gives it a model and the third opens a chat.

Five Agents Worth Creating First, Rated

Agent The one job Difficulty Verdict
Notes agent It turns a transcript into notes and action points. Easy It's the best very first agent.
Research agent It answers one question with a short summary and sources. Easy It teaches you about web access.
Keyword scout It names three keywords from the queries you paste in. Easy It's the best first agent for SEO.
Draft-reply agent It writes replies that you send yourself. Medium It teaches you about guardrails.
Daily recap agent It sends a to-do recap every morning. Medium It teaches you about schedules.

The daily recap is a real Hermes example from my guides.

hermes cron create "0 6 * * *" "Read my task list and send me a short to-do recap for today" --name "Daily 6AM To-Do Recap" --deliver telegram

If you'd like a second pair of eyes on your own business, you can book a free SEO strategy session and we'll map out your plan together. If you're already serious about SEO and want to work with other experienced operators, have a look at the SEO Elite Circle.

How To Practise: Create, Break, Fix

You create the agent from the brief in one sitting.

You then break it on purpose.

  • You give it a vague request and check whether it asks or guesses.
  • You ask for something outside its job and check whether it refuses.
  • You swap the brain for a smaller model and re-run the five tests.
  • You lower the reasoning level and compare the answers.
  • You remove a permission and see how it copes.

You then fix one thing at a time and test again.

Every failure points to a line in the brief that needs rewriting.

Free Courses If You Want Theory As Well

You don't need a course to create your first agent, but some people like structure.

I checked both of these on the providers' own pages today.

  • The Hugging Face AI Agents Course describes itself as a free course that takes you from beginner to expert in understanding, using and building AI agents.
  • Microsoft's AI Agents for Beginners is a set of lessons on GitHub that covers the fundamentals of building AI agents.

I haven't taken either one, so I can't give you a verdict on them.

My own free option is the Agent OS course video below.

If you'd like a ranked comparison, read my post on the best AI agent course for beginners.

🔥 Want a second opinion on your agent design? The AI Profit Boardroom has four coaching calls a week where you can show your agent brief and ask me and the team what we'd change. → Join the AI Profit Boardroom

Related Reading

Also On Our Network

FAQ: How To Learn Creating AI Agents

What is the best way to learn creating AI agents?

My verdict is that you should design one agent before you learn any tool.

You fill in a one-page brief, create the agent with safe settings and then test it on five real tasks.

You learn the most when you break it and fix it.

Is StarNet or Hermes Agent better for a first agent?

StarNet is better for learning, because every design choice is a visible button.

Hermes Agent is better for real work, because it supports schedules and messaging.

Both are free and open source, so you can try both.

Which design choice matters most?

The job matters most.

A clear, single outcome makes every other choice easier.

The persona matters least, although it's the most fun to set.

Do I need the best model to create a good agent?

You don't need the best model for a beginner job.

A free model or a subscription you already own is enough to learn on.

You should change the brain only when the agent fails the same test twice.

What approval level should a beginner use?

A beginner should use ask for approval.

The agent then checks with you before it acts.

You move to full power only after you've watched it work for several sessions.

Do I need a course to learn creating AI agents?

You don't need a course to create your first agent.

A brief, a free tool and five test tasks are enough.

A course helps if you want structure, and free ones exist from Hugging Face and Microsoft.

📺 Video notes + links to the tools 👉

🎥 Learn how I make these videos 👉

🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉

About Julian

I'm Julian Goldie, an SEO entrepreneur, author and founder of the AI Profit Boardroom, which has 3,400+ members.

I help business owners scale with AI agents, automation and SEO.

  • I built Goldie Agency, a 7-figure SEO and link building agency with 70+ employees.
  • I've grown a YouTube channel to 400,000+ subscribers.
  • I wrote the book "Link Building Mastery".
  • I built the Agent Operating System, which runs a crew of AI agents under one shared memory.

→ Get my best AI training inside the AI Profit Boardroom

Make the nine choices on paper, create one agent and rate your own results, and you'll know how to learn creating AI agents.