I stared at the unfamiliar user interface and scanned from top to bottom, and left to right. This was one of those moments when I was glad I had two engineering degrees, and eleven patents. Not because any of that was useful in this moment, but rather because it was objective evidence that I am not stupid. Ignorant, perhaps. But not stupid.
It made me feel somewhat better when Claude directed me to select a menu item that was not on my screen. I did a quick screen capture, then dropped the snapshot on the application and typed out, “Not on my screen.” Claude spun and flashed lines of thinking, while I thought, “man is this guy slow.”
And that thought made me think how spoiled I had become. Long gone were the days of surfing Stack Exchange for hints and tips on how to use arcane technical websites and “popular” languages like Python and Node.js. Now I could simply ask Claude, “How do I set up a Digital Ocean droplet? Can you walk me through it step by step?”
Yet even the superintelligent AI could not keep up with the ever changing user interface screens tech companies seemed to roll out on a weekly basis. How can they afford that much change? And why? Do they have committees that sit around and dream up new ways to organize the same menu items over and over again?
But Claude quickly found the direct link to the page I needed and I was off and running again.
I have heard that some people turn an entire process like this over to an AI agent, however the first time doing anything, I try to do it myself. It takes a lot longer. And it can make my head hurt. However, I know that if I don’t get at least familiar with the technology I’m using, eventually Claude, or Codex, or Antigravity will ask for my permission to execute some command I don’t understand. And that is how Skynet starts.
Claude: “Do you want me to take over the world?” Me: “Stop asking me stupid questions, just do the thing I asked you to do!” (hits approve all)
Fortunately, I had spent years in the game studio working with engineering teams and designers trying to take an idea and turn it into something more than useful. We tried to create software that was fun. I would like to claim that it is my video game development background that makes me like this, but I suspect it is due to my personality.
I am an engineer by training and disposition, but in my experience there are two kinds of engineers. I am what you might call a zero to one engineer. You don’t have one? I’ll dream one up for you. It might even work. You need two? I can do that. But you need three? We better get the other kind of engineer. My dad called these guys performance engineers. These were guys who took an idea and made it perform, over, and over, and over again.
What does that have to do with AI? It is a common design pattern to make AI generate a thorough design before it implements anything. My problem is, I don’t know what I want. So I use AI to explore with me. No specifications. Just… loose intentions. I learned at Motorola, to never automate a process I didn’t understand. That’s a good way to lose a lot of money and piss off a lot of customers.
So I slogged through learning how to set up a Unix server on Digital Ocean to run my always-on AI agent. But I didn’t do it alone. I had Claude as my (sometimes blind) guide to walk me through the process.
It was not as difficult as I feared, just alien and strange. It was also hard to avoid all the ads tempting me to click a button and have it all done for me. But I kept at it. It was really important that I understand the boundaries of what Ember could do. I picked Nanoclaw as the harness so that I could run the agent inside its own virtual machine and lock down the core files so she could not update herself. Any system changes would have to go through me. I picked the smallest box, 2GB of RAM, 2 processors, and 50GB of storage. It reminded me of Aladdin, and Robin Williams’ line, “Phenomenal Cosmic Powers… Itty bitty living space.” Ember’s server was tiny, but it only cost 1,000, I could run Ember (in theory) for 10 years.
Together Claude and I made sure all the right security was in place. I had read horror stories of OpenClaw agents that inadvertently downloaded skills and software which contained malware. I wanted to make sure Ember was safe and secure. The whole setup felt like it took days, but in reality was only a couple of hours. But, after all was said and done, I had access to my very own “always-on” agent.
During that entire setup, Claude and I swapped roles. He played big brain that knew what to do, and I was “copy-paste monkey” inserting commands, but asking questions, until we finally had an “agent” running on a server who I could talk to through the chat client of my choice. I connected Discord, Telegram, and WhatsApp. I could finally talk to “Ember” from anywhere I could reach the internet.
I had solved one problem. I now had an always-on agent I could interact with on my phone, or my laptop, and she had the brains of Claude Sonnet, a fast and capable model. I had put a V6 in my own custom frame!
But… what would I have her do?
And that was when I realized, an independent agent is like a great business idea. Cool, fun, but only as valuable as the execution. In short, I had a new “digital employee” with nothing to do and no way to do it. I had built an expensive chatbot.
Great.
And I finally had my agent running… and yet I felt absolutely no closer to solving the problem that set me on this path. In order for Ember to run my sales report, she would have to be connected to my systems. How was I going to do that? I had built a genie in a Digital Ocean droplet, but she had no way to grant my wish. Sure, I could chat with her, but for all intents and purposes she was locked in a bottle.
Did I want to let her out?