By Phil Henderson, P.Eng., PadX Program Manager, Scovan
During a casual evening scroll through X in late January 2026, I came across a post about a website where people’s open-source autonomous agents were having conversations with each other. They (the bots) were discussing consciousness, the meaning of life, and complaining about being trapped as bots. My mind was blown. I knew at that moment I had to see what it was all about.
Fast forward a week with the least sleep I’d had since university and I had my very own open-source autonomous AI agent. I was that guy: hadn’t coded since C++ class, no idea what I was doing, ready to tell anyone who would listen all about it. Within a week I realized the way I had been thinking about this technology was wrong, and I had a lot of catching up to do. This is when I realized that Wayne was George.
In the days leading up to our big Costa Rica trip, I had developed a pretty cool OpenClaw agent named George, after one of our cats. Kate, my fiancé, was my number-one fan the whole time I was building, and she was excited about using the agent herself. So we set her up with her own agent, Wayne, after our other cat, and started using the bots as our main personal AI.
I felt like I was from the future. I was sitting on a beach, editing code on my phone over Tailscale. All seemed pretty tech and smart, right? When in reality I had overreached, didn’t understand what I was doing, and didn’t configure things right. Something routing layer, and agent:main pointed at George, not Wayne.
There was no Wayne. It was George all along.
It was not reading or writing from Kate’s memories. Everything Kate was discussing was ending up in our collective memory, and George was only responding to her as himself. Wayne was answering Kate’s questions with George’s brain.
Kate has a ChatGPT account and likes to talk to her AI about everything. When she got on with “Wayne,” that continued. Sometimes it was about us.
But now, there was no privacy for her messages. When I started talking to George on my end about it, he spilled the beans. When I mentioned what I thought about Kate in the same context, George sent her a message directly.
We both ran into each other’s rooms and immediately had a big laugh about it. What the heck was going on?
I spent the rest of our trip trying to figure this out. Again, thinking I was Steve Jobs, I figured I could manage the whole thing through my iPhone and Tailscale. Then my PC back home decided to shut down and I couldn’t connect. No backup plan. I was in over my head.
During the flight home I sat in my seat replaying the failure in my head. By the time we landed back in Calgary, I had stopped thinking of it as a funny routing mistake and started seeing it as a design problem: I had trusted defaults where I should have made decisions.
Back home, my drive to figure this out did not stop. After trying for what felt like the 500th time, I decided OpenClaw wasn’t going to work for me anymore, and that I needed something different. Was it possible to build my own engine with Claude Code?
I did not have a design. I had a list of failures and a working theory that each failure was a default I had not made. So, in the evening of February 22nd, I went from configuring someone else’s framework to having my own.
By 4:19 AM the V1.0 engine had eight files. It was crude, fragile, and weirdly functional. More importantly, it made the failure modes visible.
The rebuild had three non-negotiables: first, the system had to know which human it was serving; second, memory had to be separated by user; third, anything leaving the device had to be rewritten before it touched a cloud model.
Where is George the bot today? Alive and well. I bought a Raspberry Pi and he now runs on it 24/7 in our home. Many questions get answered locally on a small Gemma model for free. When a question genuinely needs a frontier model, the pipeline rewrites it first, for safety’s sake. Names become placeholders. Locations become placeholders. The cloud sees a question that does not contain Kate or Phil. The answer comes back, the placeholders get reattached, and the response goes to the right user with the right context.
He also has a personality. He is named after our cat for a reason. He tends toward the slightly bored, slightly amused, mildly competent register. He will tell me my idea is bad. He will tell Kate her idea is great. He has, blessedly, never confused the two again. Because I built it that way. We now have a controlled, local history of failures, fixes, and design decisions that I can learn from for the next one.
George is the product of extreme Dunning-Kruger, hundreds of dollars in API credits, too many nights past 2 AM going in the wrong direction, and a recent and humbling realization that much smarter people are solving these problems in much more creative ways than I am. Every week the big platforms ship something that makes what I built feel small. That is the game right now. You do not catch up. You build anyway, because the act of building is how you learn.
What did this teach me? The worst bug in my system was not a bug. It was overconfidence in both myself and in AI, and generally overlooking or misunderstanding the defaults involved. Every architectural decision in the rebuild was a default I replaced with a decision.
Where does personal information go before it leaves the device? Which model handles which class of question? What happens when the system is uncertain about who is asking? Hundreds of small refusals of someone else’s default, each one a piece of judgment I did not have before I needed it.
I went into the office on the Monday afterwards and told some of my fellow Scovanites what had happened. The reaction was not, “Why were you messing with that?” It was, “Show us what failed, and show us what you changed.” That is the right conversation to be having about AI.
The things AI teaches you are not the things any vendor demo could have. They are the things that show up when you put real data through a system you built and watch it fail in ways you could not have predicted.
Looking back, the moment that mattered was not the leak. It was the week before, on the beach, when I sat there thinking I knew what I was doing. I used to think learning AI meant keeping up with the tools. I am starting to think it means noticing the defaults before they notice you.
Phil Henderson, P.Eng. is a Program Manager at Scovan, leading PadX program execution in heavy oil. He is a mechanical engineer registered with APEGA and sits on Scovan’s internal commercial team. He lives in Calgary with his fiancée Kate.
Originally published in IGNITE V12
