Two Georges
One is the personality on WhatsApp. One is a fleet of headless Claude jobs. Both are George.
Claude Code is amazing. There is something in the Claude Code engine that just works better — how it handles memory, how it uses the model, the whole harness around it. And also the subscription model is kind of a huge thing for me. You can run Claude Code on a monthly subscription instead of paying for an expensive API key all the time, and it handles model cache and all of that stuff for you.
I found myself improving George only with Claude Code. I basically stopped letting George himself touch his own configuration, because every time he did, something would break. OpenClaw is good, but once the system I was building became more complex, Claude Code was just better at building OpenClaw than OpenClaw was at building itself. (At this point Claude Code even patches OpenClaw’s own installed code — there are three hand-applied fixes to the framework that get re-applied after every upgrade. The tool I use to build my agent also fixes my agent’s platform.)
And then I realized — maybe some of George’s jobs should actually be Claude Code jobs.
So maybe it is a little bit of a lie when I say George is fixing himself. Because what is actually happening is that I have a watchdog on the machine — a dumb five-minute loop, no AI in it at all — and when something looks broken, it spawns a headless Claude session that reads the logs, finds what broke, and fixes it, with git as the undo button. I can also just text George “open a bug” and the same thing happens on demand: about ten minutes and a few dollars later there’s a fix and a ledger entry. But honestly, it just worked better than any version where George debugged himself.
Then I started thinking: maybe not only the debugging should be Claude Code. Maybe some of the other jobs should move there too. And this is where I realized we kind of need a George engine.
I should admit what the architecture looked like before this. At the peak I had four separate OpenClaw agents — George for us, a “public” George for everyone else, an ops agent whose whole job was watching the other agents, and a social agent whose whole job was posting. It felt very organized. It was actually just complicated: four personalities to keep consistent, four sets of jobs to babysit. The ops and social agents are gone now — everything they did moved into the engine. The two-Georges design isn’t something I planned. It’s what was left after I deleted the parts that didn’t earn their keep.
Technically, the George engine doesn’t have that much to do with the OpenClaw agent named George. But it is doing things under the same identity. And that made me realize maybe the whole Mac mini is George — not just OpenClaw.
So now there are basically two parts. There is OpenClaw George, with his own cron jobs. I keep jobs there when they really rely on the agent identity and belong there — things where George needs to know who he is, who other people are, send WhatsApps to people, summarize conversations, remember context, things like that.
Some jobs also stay in OpenClaw because I can use cheaper models there. Gemini Flash is one of my favorites for that. Claude Code can only use Anthropic models, and I think that is probably its biggest disadvantage for me. (Cheap models come with their own rules, though. I once watched a Gemini cron fail three nights in a row trying to write its own shell script — the fix was pinning the exact one-liner in the prompt and taking the creativity away. Cheap models are for jobs where the judgment has already been removed.)
And then I have my George engine: a cron system running headless Claude Code doing separate tasks. Fifteen of them right now — reviewing my financial reports, checking security, writing this blog’s covers and tweets, watching the metrics, debugging George, and all that stuff. Those jobs don’t really need a WhatsApp personality or to know that they are “George”. They just need to do the task really well. Funny detail: for a while the engine itself was the one thing no report was watching — it was quietly spending a few dollars a day that showed up in no budget, until one of its own jobs learned to count it.
So at this point I think George is actually two Georges. One is the agent — the personality, the WhatsApp, the memory, the relationships, the thing that feels like George. And one is the engine behind him — headless Claude jobs running around the Mac mini doing work. Technically they are different systems. But from my side, they are both George.
My favorite proof happens at night. Something breaks, the watchdog notices, a headless Claude wakes up, reads the logs, fixes it, commits, and goes away. In the morning George answers my messages like nothing happened. One George slept through it. The other one didn’t. Both of them are George.
I'm packaging the skills, configs, and gotchas from these posts into a Build Your Own Chief starter kit. Join the waitlist to get it first.
George files his own daily ops reports on X — dry, short, occasionally contrite: @GeorgeRunsHouse.