Agentic AI: All about agents?

Agentic AI promises personal assistants on demand – but how low is the actual barrier to entry? Our columnist Markus Sekulla runs a self-experiment: How far can you get – without a prompt engineering diploma, but with common sense – in building your own AI agent that actually gets something done?

AI agents are everywhere. Hardly a meeting happens without mentioning them. Hardly a tech magazine in the past three months without a cover story. And hardly a promise in artificial intelligence has ever sounded bigger.

But how easy is it really to build such an agent yourself? One that actually does something useful – not just spouting jokes, but taking care of real tasks?

The Buzzword Jungle

So here I am, armed with coffee, curiosity, and half a dozen tabs on “Agentic AI.” The technical terms pour in: autonomous workflows, multi-agent collaboration, self-healing systems. Impressive. And, let’s be honest: it sounds like marketing.

The more I read, the more I get a feeling we all know: people write a lot more about this topic than they actually understand. Dangerous half-knowledge, nicely packaged. Above all, I notice: I can’t find a clear, concise definition of what an AI agent actually is.

The Tool Picking Struggle: A Marketplace of Promises

Before I get to the practical part, I need to work my way through the jungle of tools and promises. The market for Agentic AI tools is a colorful patchwork:

No-code platforms like Zapier, Make, or Microsoft Power Automate advertise with “Drag & Drop your AI agent.” Tempting, but often limited to simple if-then logic.

Low-code platforms like n8n or Bubble offer more flexibility but require technical understanding.

Code-first frameworks like CrewAI, LangChain or AutoGen are designed for people with Python skills.

My strategy: test the whole spectrum.

CrewAI: Learning by Frustration

I begin with CrewAI – even though the platform is considered rather complex. But I like to dive deep into new topics, and CrewAI seemed just right for that: the platform appears in almost every second YouTube video on Agentic AI and is seen as one of the leading solutions in the field. If I want to understand how most people build AI agents, this seems like the right place – even if the entry is a bit steeper.

At CrewAI, I’m greeted with the question: “What will you build today? Let AI agents do the heavy lifting.” Oh yes, please.

What follows is… let’s call it learning by frustration.

The platform promotes a low-code approach – but after just ten minutes, I feel like I’m in an escape room without any clues. I fail at a simple question: What exactly do I have to give this agent, and what do all these terms like node, JSON, and template mean? They say building an agent is like onboarding a new employee. That’s about right.

The documentation is extensive, but somehow everyone is talking past each other. One tutorial explains “agents” and “tasks,” the next talks about “crews” and “tools.” It feels like I’m learning three different products at once.

After some time – with the help of ChatGPT – I’ve at least managed to click together an “agent” that checks in the morning, based on my calendar, whether I need an umbrella. But honestly, I can do that with a simple app. Or with Siri – for those who remember her.

The truth is: Even though platforms like CrewAI simplify a lot, the entry is just as steep as expected. At least if you're not a techie with a Python tattoo. But it still feels like, with some time investment, you could build pretty much anything here. On to the next option.

n8n – Workflow Wonder

It’s a bit easier for me to get started with the platform n8n, a kind of mind map-style automation tool with a nice drag-and-drop interface. Here I build a data pipeline: If a new article comes in via RSS reader → send to ChatGPT → generate text draft → send to my newsletter tool.

Important note: You have to switch your ChatGPT subscription to developer mode. (That costs extra, about 20 dollars per million tokens.)

The interface is more intuitive than CrewAI’s, but I still have to learn how APIs work, what webhooks do. And why my beautiful flow suddenly only returns question marks. Fortunately, the always-running ChatGPT window helps.

And still: the moment when my little DIY assistant sends me a newsletter draft independently for the first time feels a bit like when my Commodore C64 loaded a game for the first time.

Is This Already an Agent?

Technically speaking, I haven’t built an agent yet – more like clever workflow management. What’s still missing is a little loop that can analyze past newsletters and suggest improved content. Only then something like initiative comes into play – and with it, a real agent.

The line between workflow automation and “real agent behavior” is somehow clear, yet still blurry. A few more steps are needed. For example, decision logic, goal definitions, and reflection mechanisms.

Conclusion

Building your own AI on demand is already possible today, but not quite as easy as promised everywhere. Let’s be honest: without ChatGPT and YouTube, I wouldn’t have had the patience. True low-code solutions often lack structure. No-code lacks context. And for code-first, you either need time or solid prior knowledge – or a patient AI consultant in the next tab.

Text: Markus Sekulla