It's Not What You Can Do, It's What You Can Assemble

Writing my newsletter these days isn't "stare at a blank doc and grind it out" anymore.

Topic selection rotates by coverage gaps, sourcing pulls from five separate knowledge bases, the images run through one script, publishing runs through another — the part I actually do by hand has turned into "run these things in order, step in where a judgment call is needed."

One day I counted the pieces: seven or eight scripts, four or five knowledge bases, a handful of AI agents each doing their own slice. Most of what I actually do is answer one question — whose turn is it.

It stopped me for a second: by that measure, do I even "write" this thing anymore?

I used to score my own worth on simple terms: how many languages I could code in, how many nasty bugs I'd fixed, whether I could parachute into an unfamiliar system and make sense of it. Real calluses. They hurt to earn, and once earned they held up.

But that yardstick has a blind spot — it only counts what I can carry alone, not what I can pull together. A few years back I brokered a project for a friend. I didn't do much of the actual work; two or three people I knew split it between them. My part was turning their pieces into a proposal the client could read and was willing to pay for. Looking back, the money from that deal had less to do with whether I could code, and a lot more to do with whether I knew who to call and how to fit the pieces.

These two things never separated in my head — not until I started working with AI agents every day and got forced to tell them apart.

The part I still haven't gotten used to isn't that the agents do bad work. It's that my own role changed. Something used to run through me start to finish; now a lot of it gets handed off, and comes back once someone — or something — else has run with it.

What actually eats time was never "can I do it myself." It's judging what to hand off and when to step back in. It's a bit like managing a junior hire, except meaner: an agent won't complain, it'll just execute your instructions straight through to the end. Look away and it can drift a long way off course.

The cost is real too. Once you hand work off, someone has to verify it actually got done — and I tripped on exactly that. On August 10th this year I set up an automated delivery pipeline. The config had a Windows-style path hard-coded into it, but the pipeline actually ran inside a scheduled job on a Linux container that couldn't read that path. The trap was timing: on days with nothing queued to send, the program returned early — before it ever reached that config line — and logged "delivered: 0/0," which looked completely normal. That false all-clear fooled me for two straight days, until enough items had actually queued up — 27 of them — and the pipeline finally broke in the open.

Only after that did it click: verifying something actually works can't rest on "nothing threw an error." You have to make it carry a real load and watch it go through.

AI agents aren't magic. This question was already sitting there; it just got dragged into the open the last couple of years. It used to be easier to score someone purely on what they could personally do. This workflow now forces me to count what I can mobilize too — leave that out, and the score is only ever half right.

And it's not either/or. A friend of mine who freelances still writes every line himself — what he knows just runs narrow and deep enough that people can't route around him. That's a working plan too, just a different route from mine. Neither beats the other.

Next time you're sizing up your own worth, try two columns on paper: on the left, everything you can carry alone; on the right, everything you can currently mobilize and assemble into a plan. Whichever column runs dry faster is the one worth working on next.