The Agent Trap: Why I Abandoned AI Agent Workflow for a Simpler Human Driven Agent Approach
- Govind Davis
- Jan 29
- 5 min read
I honestly ended up with a completely mind-blowing experience with AI this week, and it wasn’t where I expected to find it. It forced me to think differently about what we are actually building when we talk about “agents,” and it sent me down a rabbit hole of automation that ultimately spat me back out right where I started—but with a much sharper set of tools.
We talk a lot about AI agents. Everyone wants an agent. But what is an agent, really? What are we actually building? I spent a bunch of time exploring different approaches to a very specific challenge: automating my blog content production. I record shows, I have transcripts, and I want those turned into high-quality, formatted WordPress articles without me spending hours staring at a cursor.
Where I ended up was a process that is much less “fully automated robot” and much more about deep engagement with the AI. It’s a hybrid approach, and frankly, it’s the only way to get results that don’t look like AI garbage.
The Rabbit Hole of Over-Automation
I started deep in the research phase, looking at competing approaches. If you’ve looked at the automation landscape lately, you know the players. You’ve got tools like n8n, which is huge right now. You’ve got Make (formerly Integromat), which I’ve used a lot. There are all these platforms ranging from no-code drag-and-drop interfaces to complex coding environments.
My initial thought was: *I’m going to build a fully autonomous loop.* I wanted to download the transcript, run it through a rigorous AI prompt, format it, and push it straight into WordPress.
So, I started building in Make. Theoretically, this is the dream. Make has connectors for everything. But then reality hits. I started running into these really small but meaningful technical hurdles. Specifically, the WordPress connector. To get the output to look right—to handle the structured data, the headers, the HTML formatting—it got complicated fast. I found myself thinking, *”Wait a second, am I engineering a solution or just creating a maintenance nightmare?”*
I managed to get something to publish, but I couldn’t figure out the images dynamically. I ran into issues with the structured data that WordPress expects (MCP). I was going down the rabbit hole, spending hours engineering the pipe rather than improving the water flowing through it. It was getting complicated, brittle, and the output was… fine. Just fine.
And in this game, “fine” is where your reputation goes to die. Whether it’s email deliverability—warming up your mailbox so you don’t hit spam filters—or blog content, technical reliability matters. You don’t want to be the guy sending volume with a bad reputation, and you don’t want to be the brand publishing weirdly formatted AI slop.
The Pivot: Google AI Studio and the “Pro-Code” Light
So, I scrapped the Rube Goldberg machine. I stepped back and looked at Google AI Studio.
I had tried this earlier, but I revisited it with a different mindset. Instead of trying to daisy-chain a dozen API calls, what if I just built a robust, single-purpose application using Google AI Studio?
Now, a quick caveat: The onboarding friction here is real. You’re on Google Cloud, which means you have to set up a billing account. I know, I know. It’s a turn-off. dealing with cloud billing models, old credit cards on file, and enterprise-grade complexity just to run a script is annoying. But I understand why it exists—the product complexity is massive.
Once I got through the billing hurdle, however, I loaded my script. And I literally spent just a couple of minutes tweaking the system instructions. I told it: *”Here is a transcript. I need a comprehensive, long-form article. I need it formatted for WordPress.”*
The result? It was beautiful. It wasn’t just text; it was a structured application. I built a specific interface—let’s call it the STRATTEGYS Architect—where I can simply drag and drop my transcript files. It processes them using a massive context window and spits out a 1500-word business and technical deep-dive.
This wasn’t a chat bot. It deploys as a standalone application. I could look at the code on the left—Gemini 3 Flash running the logic—and the preview on the right. It was clean. It was fast.
The 90/10 Hybrid Workflow
This brings me to the most important realization of the week. We are chasing 100% automation, but the magic happens at 90%.
When I generated the article in Google AI Studio, it was 90% there. The headers were perfect. The narrative flow was matched to my voice. But there were errors. It might get a sponsor name slightly wrong (like “Send Trucks” instead of SendCrux), or it misses a hyperlink.
If you try to automate that last 10%, you will spend 90% of your engineering time handling edge cases. Instead, I accepted the 90%.
I set up a user account for Dinah, my virtual assistant, right inside the WordPress environment.
The workflow is now:
1. AI (Me + Google Studio): Load script, generate the “Masterpiece” draft. (2 minutes)
2. Human (Dinah): Copy the output, log into WordPress, paste it in. She picks the featured image, she fixes the bad links, she checks the sponsor shout-outs, and she hits publish. (15 minutes)
This hybrid approach—where the AI does the heavy lifting of synthesis and structure, and a human applies the polish and logic check—is infinitely more scalable than a brittle fully automated agent.
Low-Code vs. Pro-Code: The Enterprise Strategy
This experience maps perfectly onto the broader strategic conversation happening in IT right now. We are seeing a divergence between Low-Code and Pro-Code approaches to AI automation.
As you can see in the landscape analysis:
* The Low-Code Approach (Democratizing Automation): This is what I tried with Make. It’s great for business users and rapid prototyping. It empowers non-technical teams to build workflows. But, as I found, you hit a ceiling with complex integrations and proprietary logic.
* The Pro-Code Approach (Specialized Engineering): This is where I ended up with Google AI Studio. You have full control over the logic. You can use SDKs, you have better debugging, and you can handle complex reasoning tasks (like writing a nuanced article) much better.
But the winner isn’t one or the other. It’s the Hybrid Model.
Successful organizations in 2026 (and today) are going to be the ones that combine these paths. You use the pro-code environment to build the “brain”—the heavy reasoning engine that processes the transcript—and you might use low-code tools or human workflows for the distribution and governance.
Conclusion: Build Your Little Machines
What I built this week is a little content production machine. It’s a micro-app. It’s not trying to solve general artificial intelligence; it’s trying to solve *my* problem of turning video into text.
It completely blew my mind because it wasn’t about the complexity of the agent; it was about the quality of the synthesis. I didn’t need agents talking to agents. I needed a great model (Gemini) wrapped in a structured interface, handed off to a capable human.
So, if you’re staring at n8n or Make and tearing your hair out because the API connector isn’t working, maybe stop. Look at the Pro-Code tools. Look at Google AI Studio. Can you build a single-purpose tool that gets you 90% of the way there?
Because getting 80-90% of the way to a masterpiece and having a human finish it is a hell of a lot better than getting 100% of the way to automated garbage. Warm up your mailboxes, clean up your data, and start designing these little productive machines. It’s worth the build.
