Who answers when the AI does the work?

(This is a story about Agent Fang, a live app — you can open it here.)

In the last edition I built Agent Fang as a flow you could read top to bottom, like a recipe — and left a question hanging: who keeps rewriting that recipe as the work changes? Here is the honest answer — mostly not me, and mostly not by hand.

It crept up on me one rung at a time. I asked an AI assistant to build a small piece of Agent Fang — the part that checks and cleans what a user types. It did. So I asked for the part that runs the steps in order; it built that too. I kept the list of steps for myself, the bit I assumed still needed a human — until I watched an assistant write a better one in a minute. So I climbed a rung: I stopped building the pieces and started writing the instructions that tell the assistants how to build them. Then the strange rung. The assistants began writing the instructions for each other. Now I talk to a single manager assistant. It reads what I want, writes the briefs, and hands them out to the others — one playing developer, one tester, one administrator, one front-end designer. These days I barely write a line of Agent Fang's code myself.

For sixteen years, the craft I was proud of was half the typing and half something harder to name — a set of eyes you grow over a decade, the kind that can find the one wrong number in a long log or a spreadsheet too big to read. The typing is gone now. The eyes are not. If anything, they work harder, because I spend the whole day judging work I did not write. It feels like a promotion: I went from doing the work to directing it. But I want to be careful with that word, because it is where people start to assume the job got cheaper. It did not. I am still a doer — developers always are — the doing just moved up a level, from writing the lines to deciding whether the lines are right.

If you run a business, here is why this matters. For years all the weight sat on the solution end — building the thing was the slow, expensive part. That part just got more efficient. So the bottleneck moves to the other end: knowing exactly what you want built. That is the reason I started this newsletter in the first place — the gap was never the technology; it was turning a real need into something a machine can be told to do. That end is your ground. And the same shift reaches into it: a manager assistant can lay out the project path, draft the instructions, even take a first pass at the work — expanding the part of the job that used to be purely human. The custom tool that was once a months-long project becomes a few days of work, and because it is built the way I showed you earlier — a plain, legible flow, not a black box — it keeps changing as your business changes.

One part of the job did not move at all. When something goes wrong — a number looks off, a customer complains, someone asks why the system did what it did — a person has to answer. You cannot get that answer from the agent. Ask a model why it did something and you get a confident story, not a reliable one. So a business needs a human who understands the system well enough to stand behind the explanation. That turns out to be the oldest part of my job: for years, my actual work was answering technical questions as a senior developer. That part did not get automated. It got more important. Take it away and you are back where this newsletter started — a tool nobody can answer for, so nobody trusts it, so nobody uses it.

Two honest things before you get excited. This concentrates the work into fewer, more experienced people — a real shift in who gets hired, and I am not going to pretend it is painless. And directing well is its own skill: a vague instruction now produces bad software faster than ever. "We can do more" is true. But the new job — judgment, at the scale of a whole team — is not the easy version of the old one. It is the harder version.

So the question I will leave you with is the one that remains when getting it built is no longer the hard part: what would you build, if the only thing between you and the tool were knowing exactly what you need?

And if you want to find out, the door is open. The setup behind Agent Fang is not a one-off — the same machinery is ready to be pointed at another project's code and put to work, yours included. The difference from handing your work to a black box is the thing I keep coming back to: a human stays in the chair, able to answer for what it built. If you'd like to see it on something real of your own, that is a conversation, not a sales call.

We're planning an RV trip to Yosemite this summer — we've camped there a few times, but never in an RV. I'll let you know how it goes.