I Built An AI Agent. It was Both Amazing And Underwhelming

Good news, kind of. I think it’ll be a while before AI starts taking over swaths of jobs and displacing people, or at least I think it’s true at the small-business level.

I've been playing around with different AI tools recently, trying to figure out ways to incorporate them into my business, and it's been an interesting adventure. On the one hand, there are things AI can do that are helpful, and I think most of us have experience with: drafting messages, summarizing large documents, and creating cool images (like above). It's genuinely useful but not transformational.

So I've started experimenting with more agentic types of processes, specifically using lindy.ai, and the experience has been interesting. The basic idea is that you can create a workflow that's based on a trigger. It could be off a timer, an incoming message, a webhook, or whatever. If you’ve ever used Apple Shortcuts, it’s the same idea. Here’s a simple example:

Every morning, my team discusses ongoing maintenance and tenant issues, and then we create a plan for the day. We record those meetings, transcribe them, and save them in a Google Drive folder.

Now, every Monday morning at 9am, it grabs all the new transcripts from the past week (that’s the purple loop) and compares them against our existing knowledge base (KB) articles. In the final blue-circled-magical-AI step, it looks for gaps and missing information and sends me an email with suggested article topics.

At this step, it's not fully written-out articles; it's just internal and external topic ideas. Eventually, Lindy will write full articles on approved topics, let me edit them, and save them to our KB. The KB, by the way, is what we use to automate message replies (therefore, in the long-run, saving time while keeping a consistent communication style).

But here’s the deal: this “simple” flow took me an hour to build. And really, it started with a prompt to Lindy’s Agent Builder, which took a first (second, third, fourth,...) pass. I had to enter some manual data (like my KB’s API Token), but most of the time I copied/pasted errors from a test run and the Agent Builder to try to fix them.

At one point, the Agent Builder told me:

“I need to stop and be honest with you. We've been going in circles for a while, and the core issue is clear... [reasons why it’s not working] .... The real fix is a different architecture. Instead of a separate LLM node, the analysis should happen inside an AI State node — which DOES have context accrual and can see all upstream data naturally.”

You see, that magical blue box used to be four separate steps that didn’t work. Which, for the record, the Agent Builder created! This current version works, but it was a process to get there, and it’s still not what I fully want...

AI Agents Are Both Amazing And Underwhelming

On the one hand, creating this was amazing because I just typed in a prompt for what I wanted, and it made it (eventually). Without the Agent Builder, I would have struggled to even get started.

On the other hand, it's a strange combination of doing something fairly simple, but in an overly complex way. It reminds me of Adobe’s Dreamweaver in the early 2000’s. Sure, it was a drag-and-drop website builder (easier to make vs coding), but in an effort to match your design exactly, it fixed everything in place with a ridiculous number of nested elements. It easily doubled or tripled the file size of a “human-made” version while being less responsive (yikes!).

I get it — there’s a balance. Dreamweaver enabled more people to build websites, but those websites were fragile. Lindy feels the same way (and Make, and n8n). It enables normal people — like me — to automate workflows, but it’s not as robust or flexible as writing code in my own server environment.

And to be clear: this is just an automation; I wouldn’t call this an “AI agent” (I mean... I will, because it sounds cooler, but strictly speaking, I don’t think it is). There’s no feedback, learning, or figuring out its own path. It’s funny, the Agent Builder feels like AI, but this output doesn’t.

I’m fairly certain it’s worth my time to create these automations, but it’s not an obvious win. It’ll partly depend on the maintenance requirements, but it’ll mostly depend on how close it gets to my quality requirements. Is it worth the eventual productivity gains if it’s only 80% as good? What about 95%? It probably depends on the process.

That’s Why AI Isn’t Taking Your Job Anytime Soon

For these reasons — it’s hard to implement/upkeep and fragile — I don’t see AI taking thousands of jobs this year. Instead, I think it’ll continue to be a solid multiplier on existing efforts. And let’s be honest, no healthy company is going to say, “let’s just do the same with less.” Instead, they’ll use the efficiencies to tackle new opportunities (because if they don’t, their competitors will). That’s what happened with calculators, personal computers, and smartphones.

Over time, automation tools will get easier and more robust, just as Squarespace does today with drag-and-drop responsive websites. But it’ll take a while to get there, and chances are good that healthy companies will adapt along the way.