The 'mistakes first' framing is underrated — most tutorials skip straight to the happy path and leave beginners confused when things break. Curious what the biggest time-sink was for you: the prompting, the tool integration, or getting the loops right?
Good question. Opposite actually. Claude Code, Cortex, OpenClaw are agents, so the basics (the loop, tool use, context, memory) become the mental model you need to use them well. Without it you hit a wall the first time one misbehaves or you want it to do something it doesn't do out of the box. I run most of my day on Claude Code and knowing what's happening under the hood is the difference between a $20 toy and a system that runs my nightshift unattended.
The Jarvis mistake is the one that costs people the most and gets written about the least. Thirty years of building things and I still have to actively fight the instinct to design the whole system before I've proved out the first piece.
The folder-as-architecture framing is the most useful thing I've read about agents in months. Strips out all the mysticism and leaves something you can actually reason about.
Yeah, I think you know pop culture is to blame, actually, because when you watch movies like Iron Man or something like that and you see this AI JARVIS that knows it all, it can do research and come back with real, cool stuff. You're like, "Okay, let's build that." The reality is we are not even close to that, and that is science fiction for now, at least. Of course Tony Stark was a genius, and not everyone is a genius.
I think you are touching the same thing I was touching on. I wanted to start big, but then I had to think about what I really want and let's start small and then build on that. That was a very good plan, because rather than going and trying to make this perfect, I just created something that works for me and then built on something that works.
It's such a strange timeline to be in lol. I started in Tech almost 30 years ago. Back then you had an idea that would take AI minutes to work through today, and you spent the first week whiteboarding it, identifying the tech stack, and organizing everything. The challenge was getting the knowledge when trying new ideas.
Sometimes I feel like I'm getting less done than 30 years ago and what I considered the hard part back then is as simple as typing into a prompt. Crazy.
Your #10 is underrated — most people default to the biggest model for everything and burn through budget fast. The real unlock is routing: GPT-4o-mini or Haiku for classification and extraction, heavy models only for reasoning. Some teams report 80%+ cost reduction just from smart routing without any quality loss.
Very informative. Reminds me of the old school cron jobs and batch files and task event schedulers. As long as the script / agent has admin access then off you go...
Great starting point for anyone getting into agents. If you want the broader AI picture beyond the technical side, fashion, entertainment, luxury, what AI actually means for different industries, check me out.
The part most people skip is defining what the agent actually needs to decide vs. what it just needs to retrieve. Building an agent that fetches data is easy. Building one that makes judgment calls — that’s where the architecture choices actually matter.
cute
That's cool! We need more simple how-tos like this. Tried to come up with my own piece of work:
https://substack.com/@larsb3/note/c-258573963?r=6zu368
The 'mistakes first' framing is underrated — most tutorials skip straight to the happy path and leave beginners confused when things break. Curious what the biggest time-sink was for you: the prompting, the tool integration, or getting the loops right?
The biggest beginner mistake is thinking an agent is “an LLM plus tools.”
That is only the visible layer.
A useful agent usually needs:
- clear task boundaries
- memory policy
- tool permissions
- failure handling
- verification
- logging
- human override
- retry logic
- cost control
Without those, the agent is not autonomous.
It is just a loop with API access.
The hard part is not making it act.
The hard part is making it act safely, cheaply, and correctly when the world gets messy.
Building your first AI agent is one of the best ways to understand what these models are actually capable of.
100%
Cool? Does this become less important with tools like Claude Code, Cortex and OpenClaw?
Good question. Opposite actually. Claude Code, Cortex, OpenClaw are agents, so the basics (the loop, tool use, context, memory) become the mental model you need to use them well. Without it you hit a wall the first time one misbehaves or you want it to do something it doesn't do out of the box. I run most of my day on Claude Code and knowing what's happening under the hood is the difference between a $20 toy and a system that runs my nightshift unattended.
Love this!
Thank you :)
The Jarvis mistake is the one that costs people the most and gets written about the least. Thirty years of building things and I still have to actively fight the instinct to design the whole system before I've proved out the first piece.
The folder-as-architecture framing is the most useful thing I've read about agents in months. Strips out all the mysticism and leaves something you can actually reason about.
Yeah, I think you know pop culture is to blame, actually, because when you watch movies like Iron Man or something like that and you see this AI JARVIS that knows it all, it can do research and come back with real, cool stuff. You're like, "Okay, let's build that." The reality is we are not even close to that, and that is science fiction for now, at least. Of course Tony Stark was a genius, and not everyone is a genius.
I think you are touching the same thing I was touching on. I wanted to start big, but then I had to think about what I really want and let's start small and then build on that. That was a very good plan, because rather than going and trying to make this perfect, I just created something that works for me and then built on something that works.
It's such a strange timeline to be in lol. I started in Tech almost 30 years ago. Back then you had an idea that would take AI minutes to work through today, and you spent the first week whiteboarding it, identifying the tech stack, and organizing everything. The challenge was getting the knowledge when trying new ideas.
Sometimes I feel like I'm getting less done than 30 years ago and what I considered the hard part back then is as simple as typing into a prompt. Crazy.
How does it access your email to summarize if you say it should have its own accounts and credentials and not give it yours?
I should have write about this more - TL:TR → Apple Script. You are not exposing any auth - but Agent is able to controll Mail App :)
Hi Pawel, love your work and this is great because I am attempting the same, so your advice is great, Declan
Thanks a lot for kind words. Yeah, thank you for reading.
Bbbbut Anankin built C3PO, not R2D2? 😢
hahaha, not intentional here!
So comprehensive! Thanks a lot.
10 thing I learn from the article:
1. Start small to avoid mistakes
2. Context is the most important constraint: make every lean for context
3. Be careful of security
4. Use git control
5. The author does not share deliverable
6. Don’t fall into the arguing trap about definition of AI agents like the author :D
7. Draft core files (.md) carefully. Update them frequently
8. Use traditional code if possible. No need AI everywhere
9. Think about AI agents as advanced automation: human has to check frequently
10. Choose wisely LLM for each task with complexity
11. Bonus: don’t blindly copy paste skills.
Your #10 is underrated — most people default to the biggest model for everything and burn through budget fast. The real unlock is routing: GPT-4o-mini or Haiku for classification and extraction, heavy models only for reasoning. Some teams report 80%+ cost reduction just from smart routing without any quality loss.
Very informative. Reminds me of the old school cron jobs and batch files and task event schedulers. As long as the script / agent has admin access then off you go...
Thanks! I was trying this to be really about BASIC stuff. As normally I am too deep into “AI Stuff”.
Very well explained👏👏
thanks, I was trying my best, as this is not my usual kind of post :D
Great starting point for anyone getting into agents. If you want the broader AI picture beyond the technical side, fashion, entertainment, luxury, what AI actually means for different industries, check me out.
The part most people skip is defining what the agent actually needs to decide vs. what it just needs to retrieve. Building an agent that fetches data is easy. Building one that makes judgment calls — that’s where the architecture choices actually matter.
Yes, that might be true. Not sure about “most people”, but I think many :)