Everyone is talking about AI agents, but they're missing something important... 🤔

Preberi v slovenščini

Hi!

Last night I was researching AI agents.

And something struck me:

Many people are currently trying to build and use AI agents, but… are we even solving the right problem?

I understand why the idea is so attractive - take for example an AI agent for answering customer questions.

You set it up once and theoretically it can answer basic questions on its own. But what happens when products change, business rules change, or a customer needs a more personal approach? That’s when it becomes clear that we need something more flexible.

And that’s exactly the problem:

Most AI agents are too rigid!

You set them up with specific tools and then they’re stuck doing exactly what you originally designed them for. Change one requirement? The automation needs to be adjusted!

Is there something more flexible?

Of course there is!

Something called the Model Context Protocol (MCP).

MCP is a protocol that gives AI models the power of conversation and tool management. It works like a “translator” between human intent and the AI system - instead of rigid instructions, it allows AI to understand what we need, which tools it can use, and how to best achieve the result.

Here’s the difference (and why I’m really excited about this):

Traditional AI Agents:

Don’t get me wrong.

They’re excellent for very specific, unchanging tasks. But as soon as you want something different, you have to upgrade the agent or create a new one.

MCP Approach:

It’s like the difference between someone who can only follow a rigid script and someone who actually understands what you need and adapts to changes in the situation.

Why a different approach? 🤔

Most people try to build automated AI agents right away.

A smarter approach is:

  1. First, let’s give AI access to the tools and data we already use
  2. Use it as an assistant in daily work
  3. When we notice recurring patterns, consider automation

It’s like giving a chat 10 extra hands - it can simultaneously access different sources and tools, while we just talk to it in natural language.

🧠 Key difference?

Instead of predetermining exactly what and how AI should work, we give it access to tools and talk to it in natural language. It knows which data it needs and how to best use them for our situation.

It’s like having a super smart assistant who has access to all our tools and can quickly find and connect all necessary information - and we just talk to them. 🤯

For example, access to:

Key difference?

These aren’t pre-programmed routines - they actually think about which information is important in each unique situation. And the best part? The community is growing fast - there are already many open-source MCP servers for different use cases!

When should you actually automate?

Don’t automate too quickly! (At least that’s what I think)

  1. Test processes manually by giving chat access to all the tools you use
  2. Monitor exactly which steps you repeat most often
  3. Only when you’ve done something at least 5 times and confirmed the pattern, think about automation

The difference between people who effectively use AI and those who don’t is growing every day. Let’s make sure you’re on the right side of this divide!

If you found this newsletter interesting, you’ll find the next one even more so, because by then I’ll have researched and tested MCPs in practice.

Check the link above to see which MCPs (connections to which tools or systems) would be useful for you and let me know (just reply to this email). I’ll check the status and how it works. 🤝

Talk to you soon,
Primoz