Ever feel like every fix reveals two more bugs, every patch needs another patch? Ever stop to ask yourself "why am I fixing all this?" instead of "what else needs fixing?"

I had to remind myself to do this on a recent pass through one of my libraries. I cut 750 lines of code and retired six previously-closed bug tickets in the process. The code that caused them was gone, and it had never been necessary in the first place. Every one of them was in machinery that had accreted around a working solution, all extensively debugged and tested, and none of it needed.

This isn't a new problem. Additive bias, the well-documented human tendency to reach for solutions that add rather than subtract, has been shaping our codebases for as long as we've been writing them. What is new is the scale. Recent research shows LLMs inherit the same bias from their training data. Public projects are increasingly closing contributions because maintainers can't keep up with the volume of AI-generated PRs.

It's about the difference between fixing a problem, understanding its root cause, and removing the conditions that make the problem possible in the first place. In the age of machine-scale addition, the third move is the one only humans can reliably choose. Every other AI-and-development talk on this schedule will offer you a way to work faster. This one offers you a discipline you needed anyway, and the understanding of why now, more than ever.

The mechanism is Decision 0: a standing default at the top of every codebase: consider what to remove before what to add. Not a rule or an ideology, but a procedural default, encoded as an ADR, an agent skill, and a PR discipline, that makes a previously invisible lapse in cultural practice observable and auditable for the first time.

An experienced developer with the discipline to delete is worth more in a codebase full of agents than three who can only add. Bring your worst legacy repo. You probably already know where to start.