VelvetShark

My AI agents journey

AI agents

This post is part of my 2025 Monthly Challenge Series.
Each month, I tackle a different area of personal or professional growth. You can find a summary of all planned challenges here.

What was the challenge?

Challenge: AI Agents Month. Spend at least 1 hour every day learning and building AI agents.

Why: Building and integrating AI agents into my daily life, starting with simple tasks and progressing to more complex automations. Starting with foundational skills and a Minimum Viable Agent (MVA) in January, this challenge will evolve throughout the year.

How well did it go?

Challenge completion rate: 100% (31 out of 31)

How hard was it?

Fairly easy, 3/10. I was already deeply interested in AI agents, and this challenge was mostly a structured way to explore my curiosity systematically.

What did I learn from it?

As I suspected, this is a very interesting subject that is worth exploring much deeper and much longer. After the initial surface-level exploration, I'm going deeper into learning not just the tools but also how those tools are built and how they work from the inside.

I want to be able not just use the AI and agents but also to tweak the underlying technology, to improve it, to adapt it to my needs.

I also learned that if something is a good hammer, not everything is a nail. There are plenty of things that can be solved with AI agents but absolutely don't have to.

More on my journey below, together with a suggested learning curriculum.

Pros

  • Another thing to learn! I'll be scared when I run out of interesting and hard things to learn.
  • Simplifying and improving my life over time.
  • Building, tinkering, improving. I love that and I use every excuse to do that.
  • Future-proofing my skillset.

Cons

  • Another thing to learn. The list of things to learn is long and only getting longer. The time to do it is scarce and only getting scarcer.
  • Whatever I learn may soon be obsolete and my learning pace might get dwarfed by self-improving AI soon.

Will I keep doing it?

Yes! I saw the value, I'm learning more of the foundational knowledge, and I'll keep building those small little helpers. At first they'll be simple and helping with the tiniest things, but slowly taking over the major parts of life. Like kids.

What's the next month's challenge?

Challenge: Sleep optimization.

Why: I haven't been sleeping enough for years. I want to get better at it. I'll track sleep patterns and experiment with different sleep hygiene practices to find my optimal routine.

My AI Agent Journey: From chat companions to workflow automation

Here's a more detailed version of what happened in the first month of the challenge (and beyond).

Starting simple with Eliza

My journey into AI agents began with Eliza, an open-source AI agent framework. I wanted to build something tangible from day one, so I:

  • Connected Eliza to my Telegram account for easy mobile access
  • Contributed a couple of pull requests to Eliza's GitHub repository, improving docs or fixing bugs I encountered
  • Gave my Eliza instance a distinct personality. I named her "Deep Blue", to honor what was, in my opinion, the first instance of what we could perceive as an AI agent.
  • Based on her lore, backstory and character traits (her character files was 500+ lines long), she created an image of herself which I used as a profile picture everywhere I set her up.
  • Spent a surprising amount of time teaching her not to write in all lowercase (a quirk in the model that needed explicit instructions to override).
Deep Blue
Deep Blue in her own eyes

As I grew more comfortable, I expanded Deep Blue's capabilities:

  • Added image generation functionality
  • Deployed her to my Discord server to have another channel for interaction and automations
  • Temporarily added access to my crypto wallet and sent a test an onchain transaction as an experiment (later disabled for security concerns)
  • Implemented PDF reading capabilities, allowing for document Q&A on Discord
  • Added web browsing capabilities for real-time information

Taking a step back: Learning the fundamentals

After my initial excitement of building and deploying, I realized I was lacking deeper understanding. Most of what I had done was essentially configuration rather than true development. I decided to take a step back and focus on learning the underlying principles of LLMs and AI agents.

This foundational knowledge helped me better understand the strengths and limitations of what I was building. More importantly, it helped me recognize when an AI agent was the right solution and when it wasn't.

The email agent experiment

My next project was ambitious: creating an email management agent using Eliza. I wanted something that could sort, summarize, and respond to emails based on their content and priority. However, I quickly ran into several challenges:

  • The implementation proved more complex than anticipated
  • Results were inconsistent and often imprecise
  • Error rates were too high for something handling important communications
  • The more I reduced errors, the more it started resembling a traditional automation tool rather than an "agent"

This realization was valuable: not every task benefits from the "agency" aspect of AI. Sometimes good old-fashioned automation is more reliable and efficient. And that's what I focused on next: simple automation. Start with deterministic rules and add AI only when it is either impossible or too hard to do with simple rules.

That's how I went deep into n8n, a powerful workflow automation tool with AI capabilities.

Pivoting to automation with n8n

Taking what I learned from the email experiment, I shifted to a hybrid approach using n8n, an open-source workflow automation tool:

  • Set up a self-hosted n8n instance on Coolify, running on a Hetzner server
  • Created an email automation workflow that:
    • Uses AI to classify incoming emails by importance and topic
    • Sends Telegram notifications for urgent or important messages
    • Handles routine emails automatically without notification

This approach proved much more reliable. By limiting the AI component to just classification (where some imprecision is acceptable) and using traditional automation for the rest, I got the best of both worlds.

I started with simple classification rules and improved them over time by updating the instructions whenever misclassifications occurred. While I initially planned to implement a formal feedback mechanism to "train" the system, I found that simply refining the prompts took me quite far.

Deep Blue
A simple email workflow in n8n that only gets better with time

I'm still using this workflow on my personal email inbox and it only gets better with time. The feedback mechanism remains on my roadmap for when I hit the limits of prompt engineering, but I haven't reached that point yet. Sometimes, simpler solutions work better.

AI learning roadmap: my recommended path

For anyone looking to follow a similar journey into AI agents, here's the curriculum I followed and recommend:

1. Understanding Large Language Models

Start with Andrej Karpathy's excellent videos:

These videos provide both theoretical understanding and practical applications, striking a perfect balance for beginners.

2. Learning workflow automation with n8n

The n8n platform offers powerful automation capabilities with AI integration:

For AI-specific n8n content:

3. Advanced AI agent development

For those wanting to go deeper:

Looking ahead

As I wrap up my January challenge, I'm excited about the foundation I've built. My AI agents are still simple, but they're already saving me time and simplifying tasks. More importantly, I've gained knowledge that will serve me well as AI continues to evolve, and I'll continue learning AI at a deeper level.

The next steps on my roadmap include:

  • Building more specialized agents for specific domains
  • Implementing that feedback mechanism I mentioned earlier (if and when I hit the wall with the current system)
  • Exploring ways to connect my agents to more of my digital life
VelvetShark

by Radek Sienkiewicz

Shark footer