The Robot Government Diaries · Entry 01
The Accidental Government
My AI agents now file tickets, hold hearings, and reject each other’s work. Somewhere along the way, I accidentally built a very small government.
Everyone keeps asking which AI skills to study. My honest answer is neither a skill nor a model. It’s governance — the engineering of alignment.
The failure
We build campaign software with a team of AI agents. Version one taught us the real failure mode:
44 screens stuck between “asked for” and “actually built.” Documentation in three places, quietly disagreeing. A fifth of all effort spent on process that changed nothing.
So I stopped shipping features. I spent the next stretch on governance instead — studying how the frontier runs multi-agent systems, then hand-rolling a scheme that fits a bootstrapped budget.
That work changed my diagnosis:
AI doesn’t fail at writing code. It fails at matching human intent — and human intent often contradicts itself.
Untangling that loop, elegantly, is the actual engineering problem. Here is the shape we landed on.
The system
Version two runs on three rules:
- Git is the only memory. Intent, decisions and disputes live in writing — not in chat, not in any model’s context.
- Judgment and building are separate seats — held by models from rival labs.
- Whatever one vendor’s AI builds, the other vendor’s AI attacks. Fresh context. One mandate: break it.
Drift
One more failure mode nobody warned us about: drift. Long sessions slowly forget what’s true — the coordinator most of all.
The fix: every seat is disposable. Kill it mid-task, boot a successor from a context pack — the architecture drawing, the ledger, its seat’s rules.
Models, it turns out, hold one diagram better than thirty pages. So the packs are pictures, not essays.
The institution remembers. No single model has to.
Proof
Last night: nine adversarial review rounds. Twenty-plus real defects caught before anything shipped. One feature went four criticals, two, two, zero.
The reviewers even overturned the arbiter — the AI at the top of the hierarchy.
The honest part
Human headcount: one. But never out of the loop.
I review every new shape of the product. I watch the governance run — and when it cracks, I upgrade the rules. That is where the human hours go.
The AI writes the code. The human writes the constitution.
The bottleneck in AI-built software is no longer intelligence. It’s governance.
— The Robot Government Diaries, entry 1.
Next entry: the day the government’s own security fence arrested the mail service. (That was yesterday. Everyone survived.)