⭐ Get the Agent OS + join 3,400+ founders inside the AI Profit Boardroom → Join AIPB ($69/mo)

GoldieBench Blog · 9 min read

How To Learn AI Agents For Free: The $0 Checklist, The Practice Ladder, And Which Free Brain To Practise On

How to learn AI agents for free: the $0 checklist, a 5-level practice ladder, free courses, hidden costs, and how the free practice brains really score.

How To Learn AI Agents For Free: The $0 Checklist, The Practice Ladder, And Which Free Brain To Practise On — illustrated hero

How to learn AI agents for free takes five free things: a free place to ask questions, free lessons, free agent software, a free model and a small job to practise on.

You can set up all five today without a card.

An AI agent is software that takes a goal, uses tools and works through the steps for you.

It has a body, which is the agent software, and a brain, which is the model that does the thinking.

Free versions of both exist, so the only thing you spend is your time.

This guide gives you the checklist, a practice ladder to climb, the free courses I checked and the hidden costs that catch people out.

Near the end, I show how the free brains you will practise on really score on GoldieBench, because that tells you what to expect from them.

How to learn AI agents for free: the $0 checklist

Here is everything you need, and what each item costs.

What you needThe free pickCostThe catch
A place to ask questionsAI Money Lab, my free communityIt is free to join.You have to post your questions.
LessonsMy free YouTube courses, plus the free third-party courses belowThey are free to watch or read.You only learn by doing the steps.
Agent softwareHermes Agent from Nous ResearchIt is free and open source under the MIT licence.You install it from the terminal.
A hosted brainA free model on Nous Portal or OpenRouterThe free variants are priced at zero.Some are time-limited, and all can be rate limited.
A local brainA model running on your own machineIt is free, because nothing is metered.Small local models are weaker.
A practice jobOne small daily task from your own workIt costs nothing.You have to run it every day.

You do not need a new computer for any of this.

Hosted free models run on someone else's hardware, so a normal laptop is enough to start.

Step one: install the free agent

Hermes Agent is the open-source agent from Nous Research, and it is the body you will practise with.

It installs with the official script, and then a setup wizard asks you to choose a model provider.

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
source ~/.zshrc
hermes setup

You can check the install with hermes doctor, and you can change the model later with hermes model.

The provider you choose in that wizard is the only part of this stack that can ever bill you.

If you would like a screen instead of a terminal, my free Hermes Agent OS post covers Herald OS, which is a free MIT-licensed desktop that runs on Hermes.

StarNet is another free MIT-licensed option, and it shows your agents inside a pixel-art space station.

I think StarNet is fun to play with, and I do not think it has many practical use cases yet.

Step two: connect a brain that is marked free

There are three free routes, and I would set up at least two of them.

Nous Portal free models

Nous Portal is the model portal Hermes connects to by default.

On 13 October 2026 the Nous inference API listed ten models priced at zero.

Those were Solar Mini 4 free, MissingNo, Step 3.7 Flash, Step 5 Preview, LongCat 2.0, LongCat 2.5 Preview, Laguna XS 2.1, Laguna S 2.1 and two Ling 3.0 Flash variants.

Please treat that as a snapshot.

Solar Mini 4 was announced as free for two weeks, and MissingNo is free for a limited time with no end date published.

My Solar Mini 4 Hermes post and my Hermes MissingNo post show how to switch each one on.

OpenRouter free models

OpenRouter lists models with ":free" at the end of the ID.

I counted 15 through the OpenRouter API on the same day, including two Gemma 4 models, Laguna XS 2.1, Laguna S 2.1, LFM 2.5 2.6B, Inkling and several NVIDIA Nemotron models.

If the ID does not end in ":free", it is not free.

Local models

A local model runs on your own machine, so there is no meter at all.

The Hermes docs describe the Ollama route in three moves.

You pull a model with ollama pull, run hermes model, and select "Custom endpoint".

Then you enter http://localhost:11434/v1, skip the API key and type the model name.

The trade-off is quality, and I will show you real scores for that further down.

The practice ladder: five levels to climb for free

Most free courses teach theory first, and I teach the other way round.

You start with one agent doing one job, and you climb one level at a time.

LevelWhat you practiseA first job to tryYou are ready to move on when
1. ChatYou talk to the agent and check the setup works.You type the single word test.It answers quickly on a free model.
2. ToolsYou watch the agent use a tool.You ask it to read a folder and summarise the newest files.It summarises the right files.
3. ResearchYou hand it a multi-step job.You ask for the last seven days of news in your niche, with sources.You can click and check every source.
4. SkillsYou make the agent reusable.You run /learn on a topic and read the skill it writes.You can run that skill again tomorrow.
5. AutomationsYou make it run without you.You schedule a daily to-do recap.It runs on time for a week.

I use that first level on every new model, because one word proves the profile, the provider and the model all work.

When I tested Solar Mini 4 at level three, it came back fast with a breakdown and a list of sources.

When I tested MissingNo at level four, the /learn command worked really well and it was quick.

A coding task was weaker, because a simple ping-pong style game took about five minutes and was not frontier level.

In my own setup, level five looks like a daily 6am to-do recap and a morning inbox summary.

Keep one profile per model

This is the habit that makes free models practical.

I create a separate Hermes profile for each model, and I name the profile after the model.

hermes profile create solar-mini-4
hermes -p solar-mini-4

When one free model is rate limited, I open another profile and keep climbing the ladder.

Free lessons I would use

My own free picks are AI Money Lab and two free YouTube courses.

AI Money Lab had more than 93,000 members when I checked, and its about page lists free AI courses, 50+ free AI tools, 200+ ChatGPT SEO prompts and 1,000+ n8n workflows.

"How to Build Agent OS - FREE Course!" is the video at the top of this page.

My two-hour "Hermes FULL COURSE" is free on my YouTube channel too.

I also read four third-party course pages on 13 October 2026.

Free courseProviderWhat the page showed
AI Agents CourseHugging FaceThe page calls it a free course with a free certification process, and it asks for basic Python.
AI Agents for BeginnersMicrosoftIt is an open GitHub repo with 18 lessons under the MIT licence.
Introduction to LangGraphLangChain AcademyThe page says "Enroll for free" and lists 55 lessons with 6 hours of video.
Introduction to Model Context ProtocolAnthropic AcademyThe registration button is labelled free, and it teaches MCP in Python.

I have not taken those four courses myself, so I am only reporting what their own pages show.

My best free AI agent course post ranks the free options if you want a list.

Hidden costs to avoid

A free stack leaks money in a few predictable places.

Hidden costWhat happensHow to avoid it
The paid model variantSolar Mini 4 is listed twice, and one entry is paid.You always pick the variant marked free.
A free window endingA model that was free last week becomes paid or disappears.You check the free list before a long task.
Paid live voiceThe Herald OS docs list the live voice engine at $0.05 per minute through an OpenAI key.You stay on the free voice engine, which runs on your device.
Topping up by habitA rate limit hits and you add credits.You switch to another free profile.
New hardwareYou buy a machine to run bigger local models.You practise on hosted free models first.

The first row is the warning from my Solar Mini 4 video.

On 13 October 2026 the paid entry, upstage/solar-mini4, was priced at $0.05 per million input tokens and $0.20 per million output tokens.

The free entry, upstage/solar-mini4:free, was priced at zero.

To check which one you are on, run hermes dashboard, click Models, click "Set main model" and read the label.

I also cannot see what Hermes Cloud hosting will cost on your account, so read what Nous Portal shows before you deploy a cloud agent.

What will eventually cost money, and when it is worth it

The free stack has a ceiling, and you will feel it at level five.

A stronger paid model is worth it when a job you run every week keeps failing on a free brain.

Coaching is worth it when you have been stuck on the same problem for a week.

Neither is worth it before you have climbed the first three levels.

Ready to go further? (paid)

This section is a paid upsell, and I am labelling it clearly.

The AI Profit Boardroom is my paid community, with 3,400+ members and four coaching calls a week.

It includes my installable Agent OS and courses on Hermes Agent, Grok Bot, DeepSeek Harness, Claude Opus 5.5 and Jev AI.

FatRank named it the #1 AI community after testing 27 communities.

I would join once you have run agents on the free stack and you want to move faster.

Which free brain should you practise on? The bench view

This is the part only a bench site can answer.

GoldieBench scores AI models from 0 to 10 on one-shot builds, where a model gets a single prompt and has to produce a working result.

That is a harder test than basic agent work, so read these numbers as a guide to raw building power and not as a grade for chat or research.

I will be plain about two gaps first.

Solar Mini 4 and MissingNo are not on the board, so I have no score for either and I will not guess one.

Here are the free brains that are on the board, with the average each one shows today.

Free brainHow you get it freeAverage scoreScored tasksWhat that means for learning
Qwable 5 27B CoderIt runs locally through MLX on a Mac.7.1441It is the strongest free local builder on the board, and it needs a lot of memory.
Agents-A1It runs locally.4.8345It can finish simpler builds and it struggles with harder ones.
Gemma 4 12B on MLXIt runs locally.3.9842It is fine for practice and weak at one-shot builds.
Laguna XS 2.1It has a free tier on OpenRouter and on Nous Portal.3.9342It is a hosted free brain for learning the agent loop, not for serious builds.
Qwythos 9BIt runs locally.2.9842It is small enough for modest machines and it fails most builds.

For contrast, Claude Opus 5 is a paid model, and it averages 8.27 across 50 scored tasks.

That gap is the honest picture of free versus paid brains for building.

LongCat-2.0 shows an 8.12 average, but that is across only 4 scored tasks, so I would not lean on that number yet.

What I take from those scores

A free brain is good enough to learn every level of the practice ladder.

A free brain is usually not good enough to one-shot a polished app.

So I would practise chat, tools, research, skills and automations on a free model.

I would move to a stronger model only when the job itself needs more building power.

The full local picture is on the local models page, and my best free AI model for Hermes Agent post goes deeper on the free picks.

If you want the wider roadmap, read how to learn AI agents.

If you are starting with zero experience, read how to learn AI agents from scratch.

Also On Our Network

Climb one level at a time on a brain marked free, and that is how to learn AI agents for free.

FAQ

How do I learn AI agents for free?

Join a free community such as AI Money Lab, install the free and open-source Hermes Agent, connect a free model from Nous Portal or OpenRouter, and practise one small job a day. Add a local model when you want practice with no rate limits.

Which free model should I practise on?

Any model marked free on Nous Portal or OpenRouter is fine for learning the agent loop. On GoldieBench, the free local model Qwable 5 27B Coder averages 7.14 and the free hosted Laguna XS 2.1 averages 3.93, so free brains vary a lot for one-shot builds.

Are Solar Mini 4 and MissingNo on GoldieBench?

No. Neither model is on the board, so there is no GoldieBench score for them. Both were priced at zero on the Nous inference API on 13 October 2026, and both are free for a limited time.

What is the biggest hidden cost?

Picking the paid variant of a model by accident. Solar Mini 4 is listed twice on Nous Portal, and the entry without the free tag was priced at $0.05 per million input tokens and $0.20 per million output tokens on 13 October 2026.

Which free AI agent courses are real?

On 13 October 2026 the Hugging Face AI Agents Course, Microsoft's AI Agents for Beginners repo, LangChain Academy's Introduction to LangGraph and Anthropic Academy's Introduction to Model Context Protocol all showed free access on their own pages.

When is it worth paying?

It is worth paying when an agent already does a real job for you every week and the free brain keeps failing at it, or when you want live coaching and a ready-made Agent OS.

The same stack Julian uses

Run this stack yourself.

Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,400+ founders shipping with it every day all live inside the AI Profit Boardroom.

3,400+founders
258documented wins
38countries
$69/momonthly