How to learn building AI agents comes down to one rule, which is that you build seven small agents in order and let each one teach you a single concept.
The seven builds are a scheduled daily recap, an inbox triage agent, a news radar, a second brain, a skill taught from a book, a kanban build line and your own agent OS dashboard.
I've published every one of these builds with real receipts, so this page is the operator walkthrough with the exact commands.
Each rung tells you what you learn, the tools, the commands, a definition of done and what usually breaks.
You don't need to code for the first six rungs, and I'll show you where the code route starts if you want it.
The 7-Build Ladder At A Glance
| Rung | What you build | The one concept it teaches | Main command or feature |
|---|---|---|---|
| 1 | A scheduled daily recap | The agent loop on a timer | hermes cron create |
| 2 | An inbox triage agent | Tools and safe permissions | The email-inbox-triage skill |
| 3 | A research and news radar | Live data and structured output | hermes -z and hermes send |
| 4 | A second brain with Obsidian | Memory and context | A vault path and an "about" page |
| 5 | A skill taught from a book or SOP | Reusable skills | /learn |
| 6 | A kanban build line | Several agents working as a team | hermes kanban |
| 7 | Your own agent OS dashboard | System design | Claude Code and plain English prompts |
Why Building Beats Studying When You Learn AI Agents
An AI agent is a model that can use tools, follow a goal and keep working without you typing every step.
You can read that sentence a hundred times and still not know how to build one.
You'll understand it the first morning a job you scheduled lands on your phone before you wake up.
That's why I teach agents as a ladder of builds and not as a syllabus.
Each rung is small enough to finish, and each one leaves you with something that keeps running.
The ladder also follows the way I teach inside my own community, which is to start with one agent doing one job and to use free models first.
Set Up Your Workshop Before Rung One
Every rung on this page uses Hermes Agent, which is the free, open-source agent from Nous Research.
You install it with the official script.
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
You run hermes setup next and connect a model, which can be a free model, a local model or a paid plan.
You run hermes doctor to check the install, and you run hermes update so the newer commands exist on your machine.
If you'd like the click-through install, my one-click Hermes setup walkthrough covers it.
The official Hermes docs are worth keeping open in a tab while you build.
Rung 1: A Scheduled Daily Recap
What you learn: You learn the basic agent loop, which is a prompt that runs, does some work and delivers a result.
Tools: You need Hermes, one connected model and a place to receive the message.
You create the job with one command.
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
The schedule accepts plain forms such as 30m and every 2h, as well as standard cron expressions.
You can see your jobs with hermes cron list, and you can read the run history with hermes cron runs.
In my Herald OS video I showed three of my own automations, which are a daily 6am to-do recap, a daily Japanese practice session and a morning inbox triage.
Definition of done: The recap arrives three mornings in a row without you opening the app.
What usually breaks: The prompt is the usual culprit, because Hermes asks for a self-contained prompt and beginners write one that assumes a chat history.
The delivery target is the second culprit, so run hermes gateway setup to connect Telegram, or use --deliver local until you have.
You should also run hermes cron status to confirm the scheduler is actually running.
🔥 Want my exact How To Learn Building 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
Rung 2: An Inbox Triage Agent
What you learn: You learn that an agent is only as useful as the tools it can reach, and only as safe as the permissions you give it.
Tools: Hermes ships a built-in skill called email-inbox-triage, and its own description says it prioritises threads and drafts replies safely.
That skill lists himalaya and google-workspace as its related mail skills, so you connect your mailbox through one of those first.
You can check what's installed with hermes skills list.
Then you attach the skill to a scheduled job.
hermes cron create "0 7 * * *" "Triage the last 24 hours of my inbox, list the five threads that need me, and draft replies without sending anything" --name "Morning Inbox Triage" --skill email-inbox-triage --deliver telegram
Definition of done: For five mornings running, the five threads it picks are the five you'd have picked, and nothing has been sent without you.
What usually breaks: The mailbox connection is the first thing to fail, and the agent will tell you it can't read anything.
Vague rules are the second problem, so write down what "urgent" means for you and put it in the prompt.
Too much permission is the third problem, which is why I'd keep the agent on drafts until you trust it.
Rung 3: A Research And News Radar
What you learn: You learn how to pull live data, how to force a structured answer and how to run an agent from a script.
Tools: You use the one-shot flag, hermes -z, which sends a single prompt and prints only the final answer.
My own radar is one hermes -z run with a strict prompt, fired by a scheduled job at 6am.
It returns six signals, and each one carries a headline, a heat score out of 100, the real post and a content angle.
The top three get pushed to my phone with hermes send.
hermes -z "Find the six hottest stories in my niche today and return a headline, a heat score out of 100, the source link and one content angle for each" --provider xai-oauth --model grok-4
Only xAI models carry the live X search tool, which is why that run is pinned to that provider and model.
My Hermes Oracle AI guide shows the receipts, which are 49 mornings logged, 294 trends scored and 46 pages published from those trends.
Definition of done: You wake up to six scored signals with real links for seven days in a row.
What usually breaks: Live search sometimes returns an empty sweep for no clear reason, and one retry on empty fixes most of those mornings.
The wrong model is the other common failure, because a model without a live search tool will happily invent the news.
One-shot mode also skips approval prompts, so keep that prompt narrow and read-only.
🔥 Want the ladder already wired, with someone to ask when you get stuck? Inside the AI Profit Boardroom, I share my Hermes automations, the memory setup and the kanban build line as step-by-step trainings. You also get four coaching calls a week and 3,400+ members building the same stack. → Get access here
Rung 4: A Second Brain With Obsidian
What you learn: You learn memory, which means giving an agent your real context so it stops answering like a stranger.
Tools: You need Obsidian, which is a free download, and Hermes has a built-in obsidian skill.
You make a new vault, hand Hermes the full folder path and ask it to build the folders with an index note in each.
You must give it the actual path, such as ~/SecondBrain, because I skipped that once while filming and Hermes built the folders in its own workspace instead.
Then you write one short "about" page that covers who you are, what the business does and how you like to sound.
My own vault held 614 notes when I measured it for the Hermes second brain guide, which is far too much to paste into a chat.
So the agent reads the "about" page first, then today's notes, then the project pages it needs.
Definition of done: You open a fresh chat, ask a question about your business, and the agent answers correctly and names the notes it read.
In my demo it listed all 11 notes it used.
What usually breaks: The missing path is the classic mistake.
Dumping everything into context is the second mistake, because one clean page beats six hundred messy ones.
Agents overwriting each other is the third, so give every agent its own folder to write into.
Rung 5: A Skill Taught From A Book Or SOP
What you learn: You learn what a skill is, which is knowledge or a procedure the agent can load when it needs it instead of carrying it all the time.
Tools: You use the /learn command inside a Hermes chat.
/learn ~/Documents/your-book-or-sop.pdf
Hermes reads the document once and writes a lean index file, a reference file per chapter, a glossary and a cheat sheet.
After that, every question goes from the index to one chapter to the answer.
I tested it on Sun Tzu's The Art of War as an epub, and the walkthrough is in my /learn guide.
The engine's own testing on real books reports 24 to 51 times fewer tokens than pasting the book into the chat.
Definition of done: In a brand new session, the agent answers three questions you already know the answers to, and it does so without you uploading the file again.
What usually breaks: An old build is the trap that caught me, because book mode only ships in newer versions, so run hermes update first.
Starting with a huge book is the other mistake, and your own onboarding SOP is a much better first document.
🚀 Ready to put How To Learn Building AI Agents to work in your business? 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
Rung 6: A Kanban Build Line With Several Agents
What you learn: You learn orchestration, which means turning a big job into tickets and letting separate workers finish them.
Tools: You use hermes kanban, and each worker is a Hermes profile with its own model and rules.
hermes profile create builder
hermes kanban boards create my-builds
hermes kanban create "Build a landing page for my newsletter" --assignee builder --max-runtime 30m
hermes kanban watch
The dispatcher runs inside the gateway now, so hermes gateway start is what gets workers picking up tickets.
In my Hermes Command Desk guide three tickets moved to done without me, which were a landing page, a content kanban app and an analytics dashboard.
Those three builds ran on free models, so my bill for them was $0.00.
Definition of done: Three tickets reach the done column, and you have opened each finished build yourself and confirmed it works.
What usually breaks: My first three dispatches all died instantly, and my kanban build line guide lists the three config traps behind that.
The first trap is a worker that boots a whole gateway and dies on a port conflict.
The second is a provider name Hermes doesn't recognise, and the third is a base URL that gets silently ignored.
Vague tickets are the everyday failure, and hermes kanban show followed by the ticket id tells you what the worker actually did.
Rung 7: Your Own Agent OS Dashboard
What you learn: You learn system design, which is how memory, agents and a task board fit into one place you actually open every day.
Tools: You use Claude Code as the builder, Obsidian as the memory, Hermes for long jobs and Tailscale for phone access.
npm install -g @anthropic-ai/claude-code
You open a folder, type claude and describe the dashboard you want in plain English.
Claude builds it and gives you a local address, which is usually localhost:3000.
Then you look at it, say what's wrong and let it fix that, and that describe, review and refine loop is the whole skill.
My free Agent OS build guide gives the exact prompts in five steps, which are the dashboard, the memory vault, the agents, the kanban board and access from anywhere.
Hermes also ships its own hermes dashboard command, which starts a web UI for config, keys and sessions if you want a head start.
Definition of done: One page shows your agents, your memory and your board, and it opens on your phone.
What usually breaks: Installing every agent on day one is the big mistake, so start with Claude Code plus one more.
Skipping the vault is the second mistake, because a dashboard without shared memory is only a prettier pile of tabs.
Automating too early is the third, and the order I teach is manual first, then assisted, then automated.
Two Tracks: The No-Code Harness Route Vs The Code Route
There are two honest ways to climb this ladder.
| Track | What you use | What you write | Pick it if |
|---|---|---|---|
| No-code harness route | A ready-made agent such as Hermes Agent | Commands, prompts and skills | You want working agents this week |
| Code route | An agents SDK inside your own program | Python or TypeScript | You want to ship agents inside a product |
A harness is the finished shell around a model, so the loop, the tools, the memory and the scheduler already exist.
An SDK gives you the parts, and you assemble the loop yourself.
These are the frameworks I confirmed are live today on their own documentation pages.
- The OpenAI Agents SDK is OpenAI's framework for building agents in code.
- The Claude Agent SDK is Anthropic's SDK, and it sits in the Claude Code docs.
- LangGraph is the agent framework from LangChain.
- The Agent Development Kit is Google's framework, and it's usually shortened to ADK.
- smolagents is the agents library from Hugging Face.
- CrewAI is a framework for crews of agents.
I teach the harness route, so I'm naming these from their own docs and not from builds I've published with each one.
My advice is to climb rungs one to five on the harness route first, because every concept carries over when you rebuild the same seven projects in code.
How To Debug An AI Agent
Debugging an agent is mostly reading what it actually did, and Hermes gives you a command for each layer.
- You run
hermes doctorfirst, because a broken install explains a lot of strange behaviour. - You run
hermes statusto see whether every component is up. - You run
hermes logs --level ERROR --since 1hto read the recent errors. - You run
hermes cron runsfor a scheduled job andhermes kanban diagnosticsfor a board. - You run
hermes sessions listto find the conversation and read the tool calls in it. - You run
hermes insightsto see token use and cost when a job feels expensive. - You run
hermes pausewhen something is misbehaving, because it stops cron and kanban dispatch until you runhermes resume.
Then you change one thing at a time, which is the part people skip.
Tighten the prompt, run it by hand, and only then put it back on a schedule.
💡 Want help climbing the ladder faster? The AI Profit Boardroom includes my installable Agent OS, courses on Hermes Agent and coaching calls where you can share your screen and get unstuck. → Join the AI Profit Boardroom
A 7-Week Plan For Learning To Build AI Agents
In week one, install Hermes and get the daily recap arriving for three mornings.
In week two, connect your mailbox and run the triage agent on drafts only.
In week three, build the radar and let it run for seven days.
In week four, make the Obsidian vault and write the "about" page.
In week five, run /learn on one SOP and test it with questions you know.
In week six, create a board and ship three tickets.
In week seven, build the dashboard that holds the first six.
Related Reading
- How to learn creating AI agents is the sibling guide in this series.
- How to learn AI agents covers the wider learning path before you build.
- The best AI agent course in 2026 ranks the courses if you'd like lessons alongside the builds.
- The best Hermes Agent use cases gives you more build ideas once the ladder is done.
Also On Our Network
- 🌐 What each of the seven agent builds replaces in a business
- 🌐 The seven builds and both tracks, scored in a verdict
- 🌐 How the seven builds become one Agent OS
- 🌐 Which model to give each build, with real bench scores
FAQ: How To Learn Building AI Agents
What is the fastest way to learn building AI agents?
The fastest way is to build small agents in order instead of watching more lessons.
You start with one scheduled job, and you add one new concept with every build after that.
Do I need to know how to code to build AI agents?
You don't need to code for the first six builds on this ladder.
Coding helps on the code route, where you use an agents SDK inside your own program.
What should my first AI agent project be?
Your first project should be a scheduled daily recap.
One hermes cron create command gives you an agent that runs every morning, and it teaches the core loop.
How long does it take to learn building AI agents?
My plan is one build a week for seven weeks.
The first three builds each fit into an evening, and the last two take longer because they join everything together.
How do I debug an AI agent that is not working?
You check the install with hermes doctor, read the errors with hermes logs and look at the run history.
Then you change one thing at a time until the run is clean.
📺 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 founded Goldie Agency, a 7-figure SEO and link-building agency based in Bangkok.
- I've grown a YouTube channel to 400,000+ subscribers.
- I wrote "Link Building Mastery", which is available on Amazon.
- I run GoldieBench and agentos.guide, where I publish my AI model tests and free Agent OS guides.
→ Get my best AI training inside the AI Profit Boardroom
Build rung one tonight, add a rung a week, and you'll know from experience how to learn building AI agents.
