The article I published on Friday took about 24 hours from trigger to done. It started Tuesday evening with a comment thread. I woke up Wednesday with the word "never" still in my head. The research ran Wednesday and Thursday early morning, and the article was finished before work. This post is the synthesis. I want to tell you how it was built, because the process is the argument.
It started with a comment thread. A CTO said he would never let an agent work without inspection. The word "never" triggered something. Alistair Cockburn did it to me: he once replied to a question I asked with a long-form article instead of a thread response. I decided to take a leaf out of his book.
The first research brief went via the Science Officer layer of my multi-agent setup because the topic touched failure mode literature that needed formal citation verification. That round came back clean. As the article took shape, a second brief emerged from the drafting itself: the invisible reasoning failure mode needed grounding in recent news. I produced that brief using my research skill, issued it to four external models, and asked specifically for high-profile incidents that made the news. When the responses came back, I could tell the first version was scoped too tightly toward academic literature, as that skill is focused on scientific work. I iterated on the brief, widened it to include news sources, and ran it again.
The PocketOS case came back from that second round. Nine seconds, a production database, the agent's own confession. That was not in my head before I ran the brief.
Then the EU AI Act angle emerged. I knew the Act classified HR AI systems as high-risk. I did not know the precise Article 14 framing for what human oversight actually requires structurally. That is just one paragraph of one article, one small piece of a much larger research programme. If you count the items in my project knowledge base that I have amassed, then this paragraph is barely a blip.
Then we iterated on the article itself. The first draft scoped the argument to software development. The reframe came from asking: what happens when you extend this beyond engineering? Every AI interaction is already a principal-agent relationship. That changed the article entirely.
The question I keep coming back to is not how much faster this was. It is whether it would have happened at all.
That would never have happened. Not slower. Not at all.
And this is a part-time hobby.
This is not a productivity story. It is a capability story. The research infrastructure made something possible that was not possible before. And the judgment: what to include, how to frame it, when the scope was wrong, when the EU AI Act detail belonged in a footnote rather than the body. That stayed with me throughout.
The agents did not write the article. They made it possible for me to write it.