The question
Does Paul use agents?
A friend asked me that this week.
I gave him a bad answer, because the honest one takes longer than a doorway allows.
Yes. But the agents are the least interesting part — and if you're deciding what to do in your own organisation, they're the wrong thing to look at first.
Two numbers
The gap between these is the whole argument
1,369
days meddling — since I joined ChatGPT, 8 December 2022. Prompts, curiosity, the odd useful afternoon.
249
days of serious use — 2026 so far. A published book, three more at pre-production, a live estate, three papers, a named team.
It isn't that the tools changed in 2026 — they did, but that isn't it. What changed is that I stopped using AI and started building the apparatus around it.
What that produced · measured 4 August 2026
One continuous record, three and a half years
I typed
1,183,741
words, by hand
Authored
10,849,614
words written
Ratio
6.54:1
machine to human
And the number that matters more than any of them: the corpus behind Mind the Gap is 3,792,153 words. I personally typed 55,693 of them. One and a half per cent.
The question you cannot answer
So — am I the author of that book?
I think I obviously am. I originated it, directed it, chose what survived and killed what didn't.
But "I think I obviously am" is not an argument. It's a feeling. And feelings do not survive contact with a publisher's lawyer, an ethics committee, a procurement panel, or a genuinely sceptical friend.
As the machine gets better, that question gets harder, not easier. Capability is becoming abundant. Evidence of human authorship is becoming scarce.
So I built the instrument instead
The Pearls — five sheets, 458 words
An audit ladder for a working relationship between a person and a machine. One principle: each stage distrusts the stage below it.
1
Ruby 1 — the self-report⚠ Still undocumented. Named as missing, not invented.
2
Ruby 2 — the deeper passWhat did you get wrong? Where did you perform certainty?
3
Independent WitnessAudits both against the raw transcript — the only primary evidence.
4
Corpus HistorianLaws, genealogy, counterfactual dependencies across every thread.
5
Final Book CalibrationThe book is ground truth. Every concept scored for survival.
6
Creative Control CalibrationOnly now do you measure the human.
Stage 6 is what turns "1.5%" from a confession into a measurement.
Conceived · mustered · managed · agentised
The agents came last, and that is the point
Conceived8 Dec 2022
I joined. Then I deleted my early chats — three weeks of the beginning are gone. Then I started keeping everything. That instinct is why there is anything to measure now.
Musteredto Jul 2026
Named workers — a book-builder, a router, an excavator. Every one a costume worn inside a single conversation. A projection, not a process.
ManagedJun–Aug 2026
The fix wasn't a better model. A war-room list. 164 memory files of standing rulings. A written constitution. A rule that nothing is ever deleted — and a boundary defining what always comes back to me.
Agentised3 Sep 2026
Only then: seven named colleagues, each a plain text file, each dispatchable. Three more run overnight on a timer.
Then I had it marked
Six of seven — and three findings I didn't enjoy
Benchmarked against Anthropic's own published reference build for exactly this. It's a tutorial, not a product — there is no button labelled Chief of Staff.
My named team did not existEight seats, duties, handoffs — written in July. When my Chief of Staff went looking, it found nothing. Prose for thirty-seven days. Not one had signed a single piece of work.
My rulebook wasn't loading135 lines of working rules, one folder below where the system reads. Months of sessions starting without it — and describing a structure that no longer stood.
My best instrument was an orphanThe audit ladder above, tagged author: unknown, three of its five sheets quietly duplicated elsewhere.
I show you that list deliberately. An audit you only quote when it flatters you isn't an audit, it's marketing.
The actual answer
Build the instrument. Then buy any model you like.
Yes, I use agents. Seven, and three more overnight. But I only built them after three years of building the other thing — a constitution, a filing discipline, a rule about deleting, and a boundary about who decides.
The order matters more than the contents. Most people buy the crew and never write the rules, then wonder why the output can't be trusted with anything that carries consequences.
Buying capability is the easy part, and it got easier again this week. The hard part — the part that still matters in five years — is answering, with evidence, the question: what did the human actually contribute here?
// Bob was here — Kilroy over the code wall