How To Learn Building AI Agents: Seven Builds In Order, And The Right Model For Each One

How to learn building AI agents is to build seven small agents in order, from the easiest to the hardest, so each one teaches you one 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.

An AI agent is a model that can use tools, follow a goal and keep working without you typing every step.

You learn that properly by shipping one, and not by reading about it.

I have built and published all seven, so the commands and receipts on this page are real.

I will give you the ladder first, then each build, then the two builder tracks and how to debug, and finally which model I would give each rung.

The seven-build ladder at a glance

What you need before you learn building AI agents

Every build below uses Hermes Agent, which is a free, open-source agent made by Nous Research.

It runs on your own machine, and you connect whichever model you like.

You install it with the official script.

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

You run hermes setup to connect a model, hermes doctor to check the install and hermes update to get the newest commands.

The method I teach is one agent doing one job, on a free model first.

Rung 1: a scheduled daily recap

This build teaches the basic loop, where a prompt runs, does some work and delivers a result.

You need Hermes, one model and somewhere to receive the message.

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 also accepts plain forms such as 30m and every 2h.

In my Herald OS video I showed three of my own jobs, which are a daily 6am to-do recap, a daily Japanese practice session and a morning inbox triage.

The definition of done is a recap that arrives three mornings in a row without you opening the app.

The prompt usually breaks first, because Hermes asks for a self-contained prompt and beginners assume a chat history.

The delivery target breaks second, and hermes gateway setup is where you connect Telegram.

Rung 2: an inbox triage agent

This build teaches tools and permissions, which decide what an agent can touch.

Hermes ships a built-in skill called email-inbox-triage, which is described as prioritising threads and drafting replies safely.

It names himalaya and google-workspace as its related mail skills, so you connect your mailbox through one of those.

You then attach the skill to a morning job with the --skill flag on hermes cron create.

The definition of done is five mornings where its five priority threads match yours and nothing was sent without you.

The mailbox connection is what usually breaks, followed by vague rules for what counts as urgent.

I would keep the agent on drafts until it has earned more permission.

Rung 3: a research and news radar

This build teaches live data, structured output and running an agent from a script.

My radar is one hermes -z run with a strict prompt, fired by a scheduled job at 6am.

It returns six signals with a headline, a heat score out of 100, the real post and a content angle.

The top three are pushed to my phone with hermes send.

Only xAI models carry the live X search tool, so I pin that run with --provider xai-oauth --model grok-4.

My Hermes Oracle AI guide shows 49 mornings logged, 294 trends scored and 46 pages published from those trends.

The definition of done is six scored signals with working links every morning for a week.

Live search sometimes returns an empty sweep, and one retry on empty fixes most of those mornings.

Rung 4: a second brain with Obsidian

This build teaches memory, which is how an agent gets your real context.

Obsidian is a free download, and Hermes has a built-in obsidian skill.

You make a vault, give Hermes the full folder path and ask it to build the folders.

You must give the real path, because I skipped it once while filming and Hermes built in its own workspace.

My vault held 614 notes when I measured it for the Hermes second brain guide, which does not fit in a chat window.

So the agent reads one short about page first, and in the demo it listed the 11 notes it used.

The definition of done is a fresh chat that answers a question about your business and names the notes it read.

Dumping every document into the vault is the usual mistake.

Rung 5: a skill taught from a book or SOP

This build teaches skills, which are knowledge the agent loads only when it needs it.

You type /learn in a Hermes chat, followed by the path to a PDF or an epub.

Hermes reads it once and writes a lean index, a file per chapter, a glossary and a cheat sheet.

I tested it on The Art of War 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.

The definition of done is a new session that answers three questions you know the answers to, with no upload.

Book mode only ships in newer builds, and that caught me out, so update first.

Rung 6: a kanban build line with several agents

This build teaches orchestration, where a big job becomes tickets that separate workers finish.

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

The dispatcher runs inside the gateway, so hermes gateway start is what gets workers picking up tickets.

In my Hermes Command Desk guide three tickets were finished without me, which were a landing page, a content kanban app and an analytics dashboard.

Those builds ran on free models, so they cost me $0.00.

The definition of done is three tickets in the done column that you have opened and checked yourself.

My first three dispatches died instantly, and my kanban build line guide documents the three config traps behind that.

Rung 7: your own agent OS dashboard

This build teaches system design, which is how memory, agents and a task board share one home.

You install Claude Code, describe the dashboard in plain English and let it build.

npm install -g @anthropic-ai/claude-code

You then review the result, say what is wrong and let Claude refine it, and that loop is the whole method.

My free Agent OS guide gives the exact prompts in five steps, from the dashboard to phone access through Tailscale.

The definition of done is one page that shows your agents, your memory and your board, and opens on your phone.

Installing every agent on day one is the usual mistake, so start with Claude Code plus one more.

The order I teach for any task is manual first, then assisted, then automated.

Two builder tracks: the no-code harness route and the code route

A harness gives you the loop, the tools, the memory and the scheduler on day one.

An SDK gives you the parts, and you assemble the loop yourself.

These are the frameworks I confirmed have live documentation today.

• The OpenAI Agents SDK comes from OpenAI.
• The Claude Agent SDK comes from Anthropic.
• LangGraph comes from LangChain.
• The Agent Development Kit comes from Google.
• smolagents comes from Hugging Face.
• CrewAI is built around crews of agents.

I teach the harness route, so I am pointing you to those docs and not claiming published builds with each one.

GoldieBench does not score frameworks, so there is no bench number for any of them.

How to debug an AI agent

After the checklist, you change one thing and run the job by hand before you schedule it again.

Swapping the model is a fair test as well, because a weak model can fail a job that a stronger one finishes.

A seven-week plan

In week one, install Hermes and get the daily recap arriving.

In week two, run the inbox triage agent on drafts only.

In week three, build the radar and read it for seven days.

In week four, make the vault and write the about page.

In week five, turn one SOP into a skill.

In week six, ship three tickets on a board.

In week seven, build the dashboard that holds the first six.

Which model to give each rung of the ladder

Now we reach the part that GoldieBench exists for.

Hermes Agent does not ship its own model, so every build above runs on whichever brain you connect.

GoldieBench scores models on one-shot build tasks, where a model gets one prompt and has to produce a working build.

That makes the scores most relevant to rungs six and seven, and a useful guide to general capability for the rest.

The numbers below are the live averages on 13 October 2026, and each name links to its model page.

Claude Opus 5 is the strongest model in this table at 8.27, which is why I would save it for the dashboard and the hardest tickets.

Qwen 3.8 at 8.10 and MiniMax M3 at 7.97 are close behind, and either one is strong enough to be your main brain.

The local models score lower on builds, but a morning recap is a light job, so a free local model can handle rung one.

That fits how I teach it, which is free models first and a stronger model only when a rung needs one.

The three kanban builds in my Command Desk guide ran on free models, so the cheap route goes further than people expect.

The honest gap is that GoldieBench does not score scheduling, email triage, memory or skills, so there is no bench number for rungs one to five themselves.

The model behind my radar is not in this table either, because that job needs live X search and not build quality.

You can pin a model per job with the --model flag on hermes cron create and hermes kanban create, so one install can use several brains.

The methodology page explains how every build is scored, and the compare page puts any two models side by side.

For local options, the local models board and the best local model for Hermes Agent go deeper.

Where to go next

How to learn creating AI agents is the sibling guide in this series, and how to learn AI agents covers the wider path. The best AI agent course in 2026 is my ranked list if you want lessons alongside the builds. The best Hermes Agent use cases gives you more build ideas once the ladder is done.

Also On Our Network

• Agent Operatorsevery command for the seven agent builds
• AI Income Deskwhat each agent build takes off a business owner's desk
• AI Tool Verdictthe seven builds and both tracks, scored in a verdict
• agentos.guidehow the seven builds become one Agent OS
• agentos.guidethe free Agent OS build guide on agentos.guide

Build rung one tonight, match each rung to the right model, and you will know from experience how to learn building AI agents.