In my last post I wrote about what AI agents might make possible. This one is about what they actually did today, which is a more useful question and, as it turns out, a less dramatic one. Nobody built a business empire from a recliner this afternoon. What happened was quieter: a handful of ordinary tasks got finished, checked, and put away, and I spent most of my time deciding things rather than doing them. That is closer to the point than any grand prediction.
The office, such as it is, consists of a chair, a laptop, a view of the hills turning color, and three colleagues. Clara is ChatGPT, and she is my partner for research, writing, and thinking things through. Dex works through Hark and handles the practical side, the part that involves logging into things, filling in forms, and carrying a task from start to finish. Finn is meant to look after personal matters. I direct the work, supply the experience, and decide whether the results are any good. That last job turns out to be the important one.
Writing the Rules Down
This morning I wrote what I'm calling our office working agreement. It's a plain document, and most of it would be familiar to anyone who has ever managed people. Results come before conversation. Research thoroughly before declaring that something can't be done. When a task is assigned and authorized, do it rather than explaining how I might do it myself. Keep track of what has been finished and what hasn't. Don't make me repeat information I've already supplied. And above all, verify everything.
That last rule deserves some explanation. A task is not complete because an assistant attempted it, or because it reports success in a confident tone. It is complete when there's evidence: a saved change, a scheduled post that shows up as scheduled, a message that was demonstrably sent. I've asked my colleagues to sort their reports into categories: completed and verified, completed but not independently verified, partially completed, blocked by a specific limitation, or not attempted. It sounds bureaucratic, but it's the difference between knowing where things stand and hoping you do.
The Detour Around Facebook
A good example came from social media. I manage a number of Facebook Pages for my various projects, and Facebook has a habit of treating any sign-in from an unfamiliar computer as an intrusion. Dex tried more than once to work through Meta's own business tools and was stopped each time by security checks that no amount of persistence would get past. Meta's developer route had its own circular verification problem.
Rather than keep pushing on a locked door, we went around it. I already had an unused Buffer account, a scheduling service that connects to Facebook Pages. I connected three Pages myself from my own computer, which Facebook was happy to accept, and from then on Dex could prepare and schedule posts through Buffer without tripping Meta's alarms. It isn't the elegant solution I would have designed, but it works, and it took minutes instead of hours.
The first real use was a post announcing my new Verve Music track, The Space Between the Walls. I asked Dex to find the best time to post it and schedule it. The answer, based on a large published analysis of Facebook posting times, was Thursday morning at nine. The post was scheduled, and then, because of the working agreement, Dex went back and checked that Buffer actually listed it as scheduled for that time. It did.
Later in the day I wrote and published a post for the Research Page myself. A copy of the draft Dex had prepared was still sitting in Buffer, so I asked for it to be deleted to avoid an accidental duplicate. Dex deleted it and then confirmed that Buffer could no longer find it. That's a small thing, but small things done reliably are what make delegation possible.
Where Judgment Stays Human
None of this means the machines are running the show. Every public post still waits for my approval. Emails that go out under my name wait for my approval. Anything that spends money or can't be undone waits for my approval. The assistants are allowed, and expected, to carry a task through its intermediate steps without checking in at every turn, but the decisions that matter remain mine.
There's a reason for that beyond caution. An assistant can find the statistically best time to post, but it doesn't know which announcement matters more to me this week, or which subject is too personal to share, or whether a particular phrasing sounds like me. Those are judgment calls, and judgment is the part of the work I'm most qualified to supply. The arrangement works best when the machines handle the legwork and I handle the choices.
Different Colleagues, Different Habits
It has also been interesting to compare how my colleagues approach the same kind of work. Clara made the original image for my previous post, an autumn study full of instruments, monitors, and a robot pointing at a map. This afternoon I asked Dex to make a variation with me in an easy chair with a laptop instead of at a desk. The result kept Clara's composition, the hills outside the window, and the cat asleep on the blanket, and changed only what I'd asked to have changed. Neither version is better in the abstract. They reflect two different ways of starting from a description, and seeing both helps me understand what each system is good at.
The same is true of the work itself. Clara is excellent at developing an idea and talking it through. Dex is more inclined to go and do the thing, then report back. I have, on occasion, found Clara agreeing to an assignment and then not quite finishing it, without a clear explanation of why. The working agreement is partly an attempt to make that kind of gap visible, whichever assistant it comes from.
What a Good Day Looks Like
By the end of the afternoon, a song announcement was waiting to go out on Thursday, a duplicate draft had been cleaned up, a featured image had been made, and a set of rules for how my small office should operate had been written down and saved. None of it was spectacular. All of it was done, and I could see that it was done.
That, I think, is the real measure of whether this experiment is working. Not whether an AI can hold an impressive conversation, or describe what it might accomplish someday, but whether at the end of the day I can point to finished work and know it's real. Today I could. The office in the easy chair is open for business, and its first policy is a simple one: show me.
Adam Sweet Research — Independent investigations, historical research, and interesting questions.
