From Zero To SaaS #43 Learning AI Agents

Every day, I see yet another AI agent post on the top of LinkedIn with 700+ comments.

If LLMs were alcohol, the go-to relaxation drug in the Western world, then agents are what ecstasy is for millennials and the younger generation.

Aka agents are a quick gateway in business, allowing faster results for a fraction of the cost.

What/Who started the agent revolution?

The origin of Agentic Architecture is the concept of ApE prompting, Automatic Prompt Engineer, where LLM creates prompts for itself.

The genie came out of the flask when frameworks like AutoGPT ignored all safety protocols, which big AI players like Google, OpenAI, and Anthropic abide by, and allowed LLMs to execute their code using everyday language.

• Trigger: The signal that sets the agent in motion.( Slack, Email, Website Visit, etc)

• Agent: Has system prompt: natural language explaining WHO the agent is, WHAT agent jobs to be done are, and which TOOLS you use for it.

• ChatModel: To do reasoning, you can select one of the models of the LLM provider.

• Memory: The agent can store information between the use of each tool, so each tool builds upon the next.

• Tool: API call where the information is transformed or sets action in motion.

• Reaction: Agents report what they have done at the end.

What is the difference between an AI agent and an automation?

Automation doesn't use reasoning to determine the best order of actions. Like the grass mower, it can either mow the grass or stand still; its functionality is predefined.

Agents are action-aware and can create their schedule to achieve a desired outcome. It's like cooking a recipe, you know the steps, but you adjust them according to the ingredients available, cooking equipment, and time horizon.

What should start learning agents?

Anyone, especially the nontechnical, can learn agents due to platforms like Relevance AI, Make, and Zapier.

You need to understand two concepts:

• API: How platforms interact with each other.

• Workflows: How to connect modules/steps/tools to get the desired outcome

What is a realistic ROI for agents?

There are no golden bullets, so I say 10% better in terms of speed, costs, and efficiency.

My Learning Protocol

My learning protocol is based on the concepts

• You only need enough to get you going; ChatGPT can fill in the rest.

• Asking people who are further ahead how they learn or fix problems, thanks to Joep Baks and Nadia Privalikhina, for the inspiration and advice.

• Repetition can justify watching and listening to content at 2x speed.

Just In Time Learning (ChatGPT)

When people see AI, they see speed and cost reduction.

I see the teacher I always wanted but could never get.

Your curiosity is the only limit to what you can do with ChatGPT.

You can compute on the entirety of what humans have imagined, said, or written in a few seconds.

My ChatGPT Habits:

• I never met Alex Hormozi, although I consumed many of his videos. But I have regular conversations with the guy on improving my business. Just prompt:

• I am a spaghetti coder, aka sloppy, unstructured code, but I learned how to fix my and AI-generated mistakes.

• I don't know the best workflows or all the available api's

Learning Insight #1 A fatigued brain cannot see that the quality of decision-making has decreased; therefore, it looks like you are pushing forward but going backwards.

Counter: Set limits on how far you want to go that day and stop at the peak.

For me, that is 1-3 hours in, depending on the day.

Learning Insight #2: Unless you have written out your learning habits, you will never spend time learning, and you are using systems made for the general public.

Ultra Learning in 2025

Information Diets: Content consumption should be adjusted to your learning style; everyone has two.

When I found I listened 2x as fast as I read, new doors opened.

Define Learning Goals: I learn to increase the range of possibilities of my thoughts. I might not have all the life skills, but I have, on average, more perspectives, which helps cross the gaps.

Learning Reps: I believe that information isn't stored if I don't take notes. Learning is not something you make time for; it's a lifestyle philosophy. You have your phone with you, so you can always make notes.

Unlearning: It's not about how much you know or how many skills you have; it's about how fast you can remove and add skills.

Everyone is your teacher: Everything I say, I already know, so I have learned that asking quality questions helps you and the other person get more out of life.

Objectification: Life exists on the present and imaginary planes. The secondary plane is rich in past experiences and plans, but you won't capture insight without a practice like one week back and two weeks forward.

The Agent Road Map

Standart Route

Start with surface concepts, build an agent every day, and start engaging with your peers.

If you like programs, cohorts, or communities, join. Because when you pay per month, you will be incentivised to continue your learning journey.

Neurodiverse Route

I never walked the conventional route in life because I don't believe in information stacking like a ladder.

Like entrepreneurship operates with limited experience, the best learning is fuzzy and nonlinear, but you still need to make the agents.

But I notice that I am not great at learning in a group because they either are too fast or too slow.

Picking an Automation Provider

If you want a consistent UI, I recommend using an industry-standard platform like Make (most API integrations) or N8N (versatile) instead of the newer startups.

Relevance AI (more granular as N8N) and Lindy (a better version of Zapier) are the best new kids on the block.

The videos on a platform are often outdated, so basic that they have no real-life use case.

I watch most tutorials to learn one part of the flow, like connecting voice memos to an agent, storing images in the database, etc.

In a world where you can use vague intention to write code and build complex marketing funnels, it seems idiotic that an API can self-correct a JSON List into a JSON object if necessary.

For every agent you are going to build, you will learn that:

• Most platforms don't update their documentation, nor notify you if their API depreciates, aka stops working, or when they change how to use the API.

• Even if you declare the correct data type, automation platforms often change that data type for no apparent reason.

The webscraping is best done with Firecrawl, which converts the website to LLM-ready data.

I am an OpenAI fanboy, but Gemini is better for specific cases like the longest context window.

If you watch enough tutorials on code and agents, you will see that something always goes wrong.

Some rules: Keep it short, talk about what you deliver, not what it is. People don't care. If you share an automation, use Tella or Loom to record yourself.

BONUS ------> PRICING AGENTS

The truth is that the agent is worth what the other is willing to spend on it.

To get the best price, ask above your weight in cost per hour.

The client will negotiate down, and you have completed the spiel as old as time.

Senior: 130,- x per hour OR 10K + (Large and Enterprise)

Medior: 65,- x per hour OR 4-6k per agents ( Medium Sized Companies)

Junior: 40,- x per hour OR fixed cost 1-2k per agent (start-ups and small companies)

Happy Building!

If it were easy, you wouldn't find it valuable.