How To Learn AI Agents: The 8-Step Path, A 4-Week Plan, And Which Model To Learn On

How to learn AI agents is an eight-step path, and you take the steps in order until one agent does a real job for you without being asked.

1. You understand what an agent is.
2. You use one ready-made agent for a real task.
3. You give it tools.
4. You add memory.
5. You schedule it.
6. You teach it skills.
7. You run several agents.
8. You build your own agent OS.

I'm Julian Goldie, and I run a crew of AI agents inside my own Agent Operating System every day.

This post explains each step, gives you a four-week plan and lists a free resource for every step.

Near the end, I answer the question a bench reader asks next, which is which AI model you should learn on.

How to learn AI agents: the whole path in one table

The order is the lesson.

Most people who give up started at step seven, with a team of agents and no job for any of them.

I teach one agent doing one job first, and everything else is added on top of that.

Step 1: understand what an AI agent is

A chatbot answers a question, and an agent finishes a task.

Anthropic's engineering guide says agents are systems where the model dynamically directs its own process and tool usage.

It contrasts that with workflows, where models and tools follow predefined code paths.

In plain words, an agent has a brain, hands and a loop.

The brain is the AI model, and the hands are tools such as a terminal, web search and file access.

The loop means it looks at the result of each action and decides what to do next.

That brain is the part this site measures, and I come back to it near the end.

The same guide recommends the simplest solution possible, and I agree with it.

Step 2: use one ready-made agent for a real task

You learn agents by running one, so you do not start by building one.

I start people on Hermes Agent, which is a free, open-source agent from Nous Research.

The official docs offer the Hermes Desktop installer or a one-line terminal install.

After the install, hermes setup --portal logs you in and lets you pick a model.

Some models there are marked free, so check the label before you confirm.

Free models get rate limited at busy times, so hermes fallback add gives you a second one.

The built-in hermes doctor health check explains what is broken in plain words.

Then you give it one real task from your own week, such as a summary of a folder of your notes.

You watch how it plans, acts and checks its own work, and that is the moment the idea clicks.

Step 3: give it tools

An agent without tools can only talk.

The command hermes tools lists every toolset as enabled or disabled.

On my own install of version 0.21.4 it listed 26 built-in toolsets, and 20 of them were on.

The terminal, files, web and scheduling toolsets are on by default.

Outside apps connect through MCP, which the official MCP site describes as an open-source standard for connecting AI applications to external systems.

In Hermes you browse them with hermes mcp catalog.

The first mistake in my crash course is adding too many tools too soon.

You add one tool, test it on a real task and then add the next.

Step 4: add memory

A normal chatbot resets to zero every day, and memory is the fix.

Hermes ships with a small built-in memory that persists between sessions.

The official docs describe a MEMORY.md file limited to 2,200 characters and a USER.md file limited to 1,375 characters.

Past sessions are stored in a local database that the agent can search.

I add a plain notes vault on top, and I use the free Obsidian app for mine.

Three notes get you started, which are an about-me note, a decisions log and a daily log.

One rule runs it, which is that the agent reads before it works and writes before it leaves.

Memory also survives a change of model, which matters a lot once you start comparing models.

When a new model comes out, I change a dropdown and the new brain inherits the same vault.

Step 5: schedule it

A scheduled job is a prompt that runs on a timetable and delivers its result to you.

This is a real command from my use cases guide.

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

The schedule also accepts plain forms such as every 2h.

You list your jobs with hermes cron list, and hermes pause is the emergency stop.

A job that calls a model uses tokens on every run, so I'd start with one or two.

Step 6: teach it skills

A skill is a reusable instruction pack that the agent loads only when a job needs it.

In Hermes each skill is a folder with a SKILL.md file inside it.

You see what you have with hermes skills list, and every installed skill becomes a slash command.

You can correct the agent once and have that correction saved as a skill.

The /learn command turns a guide or a book into a skill.

Skills are also how a weaker model gets better results, because the method is written down for it.

Step 7: run several agents

You add a second agent when the first one has too many different jobs.

In Hermes you create profiles, such as hermes profile create research and hermes profile create builder.

Each profile has its own memory, its own skills and its own model.

My rule is one profile per model, so I can see which brain did which job.

One agent can also split a job across sub-agents, and a ticket board lets workers pick up tasks.

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.

My crash course warns against five profiles with zero workflows, and against trusting sub-agents unchecked.

Step 8: build your own agent OS

An agent OS is one dashboard where your agents live, share one memory and pick up tasks from one board.

You already have the memory from step four and the board from step seven.

The quick dashboard is hermes dashboard, which opens a web dashboard on your own machine.

The full version is one you direct, by describing it to a coding agent, reviewing the result and saying what to change.

My own Agent OS has seven agents wired in, and it is the same dashboard that built the demos on this site.

My free video course on building one is at the top of this post, and the written version is my AI Agent OS guide.

What to do each week for four weeks

This is a plan and not a promise, so repeat any week that does not end with something working.

Week four is where a bench reader has fun, because two profiles on two models let you compare them on your own work.

A free resource for every step

I checked each outside resource on its own page on 13 October 2026, and I have not taken the outside courses myself.

Two broader courses cover the theory behind the whole path.

The Hugging Face AI Agents Course describes itself as a free course that takes you from beginner to expert, with a free certification process.

The Agentic AI course from DeepLearning.AI is taught by Andrew Ng and covers reflection, tool use, planning and multi-agent workflows.

Its page says you can audit it for free, and it asks for intermediate Python skills.

LangChain Academy says its course content is always free, and it charges for its certification exam.

My free community is the AI Money Lab, which includes a free AI course and 1,000 AI agents.

My paid community is the AI Profit Boardroom, which has 3,400+ members, four coaching calls a week and my installable Agent OS.

You can also read how to learn AI agents from scratch, how to learn AI agents for free and my best AI agent course list on this site.

How long does it realistically take to learn AI agents?

It depends, and I will not invent a number for you.

Those are the only timings I can point to.

How long it takes to become good depends on the hours you give it and whether you have a real job for the agent.

Seven mistakes that slow you down

These are the seven mistakes from my Hermes crash course.

1. You add too many tools too soon.
2. You save everything to memory.
3. You put a weak model on hard work.
4. You turn every idea into a scheduled job.
5. You trust sub-agents without checking them.
6. You create five profiles and have zero workflows.
7. You treat the agent like magic, when it only compounds if you feed it well.

The third mistake is the one this site can help you with, and the next section covers it.

Which AI model should you learn AI agents on?

An agent such as Hermes does not ship with one fixed model, so you choose the brain.

Learning AI agents is a method and not a model, so no course or learning path has a GoldieBench score.

What the board can show you is how the models you might plug in perform on real one-shot builds.

GoldieBench scores each model's rendered builds from 0 to 10, across games, simulations, visuals and pages.

It does not measure tool calling, memory or scheduling, so treat these numbers as a guide to raw build quality.

These are the live averages for models that suit different stages of the path.

The pattern is clear when you read the table from top to bottom.

The free models are good enough to learn every step on this page.

The small local model shows the third mistake in action, because it scores far lower on hard builds.

My advice is to learn the method on a free model and move only the hard work to a stronger one.

That is the same mix I describe in my crash course, with a frontier brain for the hard thinking and a local model for the grunt work.

Scores move as I re-test models, so check the live model page before you choose.

The local models board ranks the models you can run on your own machine.

You can put models side by side on the compare page, and the scoring rules are on the methodology page.

My best Hermes Agent model post goes deeper on choosing the brain.

Also On Our Network

• Agent Operatorsthe eight steps to learn AI agents, with every command
• AI Income Deskhow a business owner learns AI agents around one real job
• AI Tool Verdictevery resource for learning AI agents, compared and rated
• agentos.guidehow to learn AI agents by building your own Agent OS
• agentos.guideJulian's full Hermes AI agent crash course

Install one agent, give it one real job on a free model, and add one step a week, because that is how to learn AI agents.