How To Learn More About AI Agents: Where To Keep Learning, And How To Test Every Launch Yourself

How to learn more about AI agents once you know the basics comes down to testing more and scrolling less.

I am Julian Goldie, and I run this leaderboard by testing new AI models on the same prompts almost every day.

Here is the short path.

1. You follow the release notes of the one agent you use most.
2. You watch one build video a day and copy one thing from it.
3. You join one community so you can ask when you get stuck.
4. You test every new tool and model with the same fixed prompts.
5. You turn each thing you learn into a small build within 48 hours.

The first part of this post shows where to keep learning, ranked by format.

The second part gives you a weekly plan and my launch-day test.

The last part covers how to follow models, with real scores from this board.

Where to learn more about AI agents, ranked by format

This is my own list, so my own channels come first, and I am saying so plainly.

Videos come first because watching a tool run is the fastest way to judge it.

Papers come last because they rarely change what you build this week.

Three sources are enough to start, which means one video channel, one community and one changelog.

The 30-minutes-a-day plan

Each day has one job, and the whole week takes three and a half hours.

Only two of those days are about new information.

The other five are about building, testing, asking and reviewing, because that is where learning sticks.

Friday is the day this site is built around, so I will spend the most time on it.

Official docs and changelogs

Docs and release notes are written by the people who shipped the change, so they are the most accurate source you have.

These pages move quickly.

When I checked the Hermes Agent releases page on 13 October 2026, it listed four releases in about three weeks.

Version 0.21.4 came out on 21 September, version 0.21.5 on 24 September, version 0.21.6 on 8 October and version 0.21.7 on 12 October.

One feature a week from your main tool's release notes is enough to keep you current.

GitHub repos, newsletters and X accounts

A repo shows you when a project was last updated, and every repo below had commits in October 2026 when I checked.

You set Watch to Releases only on three or four of them, and GitHub tells you when a new version ships.

For reading, I would pick one of these, and each had published in the week before 13 October 2026.

On X, I would follow official accounts that each organisation links from its own website, such as @NousResearch, @AnthropicAI and @GoogleDeepMind.

I post model launches and screenshots from this bench on my own X account.

X tells you that something launched, and the docs tell you what it actually is.

How I test a new tool the day it launches

Reading about a launch teaches you very little, and running it teaches you a lot.

My routine is on my about page, and it has three steps.

1. I wire the new model into my Agent OS, so I can send it work from the same board my other agents use.
2. I run my fixed prompt set through it with one shot each, no second tries and no hints.
3. I score each result from 0 to 10 on camera.

I score on whether it ran, how close it came to the brief and how good it looks.

Every cell on this leaderboard is one of those one-shot builds, and the methodology page explains the rules.

Each result has to be a single HTML file that runs in a browser, so anyone can open it and judge it.

How to run your own launch-day test

You do not need my setup to copy the method.

1. You save three to five prompts in a text file, and each one is a real job you care about.
2. You add the new tool or model next to the one you already use.
3. You run every prompt once and save what comes out.
4. You score each result from 0 to 10 and write one line about why.
5. You compare the scores with your current tool before you change anything.

The prompts must stay the same every time.

Fixed prompts are what turn an opinion into a comparison.

One shot matters too, because a tool that needs five retries will cost you five times the attention in real work.

After a few months you will have a small score sheet of your own, and it will know your work better than any public board does.

How to avoid hype

I use three checks, and they take about ten minutes together.

The first check is the repo and the README.

A real tool has install steps you can follow and recent updates you can see.

The second check is the date.

Anthropic's engineering post, Building effective agents, was published on 19 December 2024.

Anthropic has since added a note at the top saying much of the tooling it describes has changed.

The ideas are still worth reading, and the note is a good reminder to check the date on everything.

The third check is your own test.

A launch post will always show the tasks a model is best at.

Your prompts show the tasks you actually need.

How to turn learning into a build

My rule is that one thing learned becomes one thing built within 48 hours.

I teach the same order every time.

1. You start with one agent doing one job.
2. You keep one profile per model, so you can swap the brain without breaking the job.
3. You use free models first and pay only when a job needs it.
4. You connect the jobs into an Agent OS once two or three of them run well.

My own example is a radar.

It is an agent on a schedule that reads live posts on X at six each morning and scores every story.

When I published the guide to it, it had logged 49 mornings and scored 294 trends.

Your version can be an agent that checks three release pages each morning and sends you a summary.

If you want the full dashboard, my Agent OS build guide walks through it in five steps.

Papers for the curious

You can skip papers and still become very good with agents.

If you like knowing why things work, these well-known papers are free on arXiv.

You read the abstract and the introduction first, and you stop if it will not change a build.

How to follow AI models with a leaderboard

An agent is a model plus tools, memory and a loop, so the model you choose matters.

New models now arrive faster than anyone can read about them.

A leaderboard lets you follow them in five minutes.

GoldieBench measures one thing, which is whether a model can ship a finished build from one prompt.

It does not measure every skill an agent needs, such as long tool-calling sessions or memory.

Here are averages for some of the models on the live board on 13 October 2026.

Scores change as I add tasks, so the live pages are always the source of truth.

Three lessons the board teaches about learning

The first lesson is that newer is not automatically better.

Claude Opus 5.5 is the newer model, and it averages 7.57 here while Claude Opus 5 averages 8.27 on the same prompts.

Opus 5.5 was scored in a separate, stricter run than Opus 5, so I treat that gap as a reason to test, not as proof.

That does not make it a worse model overall, because this bench tests one kind of work.

It does mean you should test before you switch.

The second lesson is that a system can beat a single model.

Fusion and Hermes MoA both combine several models, and both sit near the top of the board.

That is a good reason to learn about orchestration and not only about models.

The third lesson is that the best-known labs are not your only option.

Qwen 3.8, Kimi K3 and GLM-5.2 all average above 7.7 here.

You can see models that run on your own machine on the local models page.

Free models first is a sensible rule for learning, because your mistakes cost nothing.

How to use this board in your Friday slot

1. You open the leaderboard and look for a model you have not tried.
2. You open its page and click two or three of the live demos.
3. You compare it with the model you use now on the compare page.
4. You add it as a new profile in your agent if it looks promising.
5. You run your own fixed prompts and write the scores down.

My scores tell you where to look.

Your scores tell you what to use.

Where to go next

If you want the path from the start, read my how to learn AI agents guide.

If you want structured lessons, my best AI agent courses online list and my best AI agent course for 2026 compare the options.

If you want to compare communities, my best Skool courses list ranks them.

If you want to know who runs this board, my Julian Goldie page covers my background.

Also On Our Network

• Agent Operatorsthe operator's weekly routine for learning more about AI agents
• AI Income Deskhow to turn AI agent learning into real builds for a business
• AI Tool Verdictnine kinds of AI agent learning source, ranked and scored
• agentos.guidethe 30-minutes-a-day learning loop inside an Agent OS
• agentos.guideJulian's full Hermes crash course on agentos.guide

Pick three sources, test one new model every Friday, write your scores down, and you will have your own answer to how to learn more about AI agents.