Is This Worth Printing?

How does agentic development go?

My feeling today is that it resembles the board game “Tsirkus” (“Circus”). You have let feelings, ideas and other oscillations write the code, and you have already come within a few steps of the finish. Then you suddenly land on a square that rides both the piece and your mood back down to the bottom.

You are disappointed and a little glad at the same time. So the human is still needed after all.

Fortuna, the patron saint of the artificial mind, was not on this repository’s side this time. The code went wobbly along with the idea, even though this was only the first or second attempt.

A game is a game. What do you do with code you have kept growing with the help of AI?

Nothing to be done. If you no longer understand it yourself and do not want to do the work required to understand it, it has to be thrown away without mercy.

There are plenty of textbook failures where, without reading the code and without understanding the decisions, people end up who knows where within a single quarter. The same happens, of course, over years with code written by humans. The difference is mostly in speed. Agentic development lets you grow confusion at a tempo one could previously only dream of.

If the generated code has not yet reached production, giving it up is relatively cheap. And yet it is human to hope that the risks will not materialise and that the software born from the interplay of your will, your prompts and chance simply works.

Unfortunately, hope is not a control mechanism.

At some point the agonies of finishing begin. You have to remove the code, the features and the skills that do not serve the goal. If neither you nor the agent understands what is going on any more, the solution has not become complex. It has simply gone out of control.

Many hand-written codebases — organic code, so to speak — probably arrive at this point as well. Somewhere along the way vigilance disappears. Careless edits are made without anyone noticing, new layers are added “just in case”, or you discover that your wishes are not in accordance with the laws of nature.

Then you have to be merciless.

Find the goal again. Write it down in a file called “GOAL.md”. Remove everything that does not serve it. If necessary, start over.

The strange thing is that sometimes you also have to write down the non-goals. Otherwise the kind model wants to heal all the pain in the world. If you do not tell it what it is not allowed to solve, it will quickly find ten more problems nobody asked it to touch.

The central skill of agentic development is no longer generating code. Code can be had in quantity, quickly and cheaply. What becomes central is deciding: what to build at all, what to believe, what to keep and what to delete.

In the first post (At the Printing House Door) the scriptorium’s scribe was startled by how cheap the reproduction of books had become. Once copying a text became easy, the main question was no longer how to make a book. The more important question became which manuscript is worth reproducing at all.

With agentic development we are in a similar place.

Producing code keeps getting cheaper. Because of that, the biggest challenge moves into the wording that precedes generation, and into the editing that follows it.

Before we send a specification to the printing house, we ought to ask:

Is this worth printing?

«“Whenever you feel an impulse to perpetrate a piece of exceptionally fine writing, obey it wholeheartedly and delete it before sending your manuscript to press. Murder your darlings.“»

Sir Arthur Quiller-Couch