Building My Fake Staff - Who Now How Build Each Other
I’m debugging three AI systems that are all under active development… at the same time.
- Autonomous Goal — my autonomous employee that wakes up and works toward long-term goals.
- “Gosh Durn It, Do What I Say” Orchestrator — the orchestration engine that makes sure LLMs actually follow instructions instead of getting creative at the wrong time. (Yes, that’s really what I call it.)
- The actual mission project they’re working on.
I’m running Codex on Linux for this.
Each layer is complex.
Each layer is changing.
Each layer can be the source of the next failure.
Eventually I got tired of playing whack-a-mole. So I gave Codex one instruction:
Run the Autonomous Goal service. If it fails, determine which layer is actually at fault. Fix it. Run again. Repeat until everything works. Then summarize what you found and fixed.
And then…I stopped typing.
For the last couple of hours it’s been doing exactly that. It has been preserving evidence, classifying failures, fixing bugs in the correct layer, rerunning the entire system, discovering the next issue, fixing that, rerunning again…
Over and over.
Sometimes the bug is in the mission.
Sometimes it’s in Autonomous Goal.
Sometimes it’s in the Orchestrator.
The important part is that it determines which one before changing anything.
The process now looks something like this:
see the diagram
The fascinating part isn’t that it fixes bugs. It’s what kinds of bugs it’s finding.
So far it has uncovered things like:
- prompt design flaws
- orchestration logic defects
- worker routing bugs
- validator edge cases
- review process inconsistencies
- environment mismatches
- playbook authoring mistakes
- selection logic issues
Every fix hardens the platform for the next mission.
This is still very much a work in progress. It’s powerful… and still fragile enough that I wouldn’t hand it production systems unsupervised.
But this feels like crossing an important line.
Instead of me debugging one bug at a time…
…I’m debugging an autonomous employee that is learning how to debug itself.
It’s equal parts exciting, terrifying, and honestly pretty funny. Ok, I threw that terrifying in not for me, but for my kids and sister when I tell them this story.
Am I replacing myself? No. I’m moving my effort to a different level of the stack. I’m spending HUGE amounts of my time, energy and creativity learning how to put all this together. AI isn’t changing “is Lee learning” - but “what Lee’s learning” and the leverage I get from having learned.
