Extreme AI Programming

For the last eighteen months, most of the code shipped by the team I run has been written by agents. That single fact has slowly reshaped every habit I thought I understood about how software actually gets made, and the gap between what Agile assumed in 1999 and what the work looks like now has become too wide to keep quiet about.

There is a harder part underneath that gap, which the industry has been slow to talk about openly. The junior job market has all but disappeared in the last year, because a senior engineer with a competent agent now ships what used to take a team of five. The obvious economic move is to hire the senior and skip the juniors, and that move is being made everywhere, quietly. Unless those of us already inside reshape around what the next generation has to bring, the industry will run on the experience it has now and stop restocking. Part of why I am writing this now, rather than waiting, is that I think we have about two years to get this right.

So I am starting a weekly essay series, which will become a short book. It is called Extreme AI Programming, the title a homage to Kent Beck's Extreme Programming Explained. The first piece publishes on Tuesday 5 May at 8am UK, with one article a week after that for roughly twenty weeks.

The argument, compressed. Agile is a coordination protocol for humans, and a good one. But the shape of the coordination problem has changed. The developer is increasingly not a person. It is an agent, often several at once, each running its own conversation with a different member of the team, each producing work on its own cadence. Standups, story points and sprint velocity do not survive contact with a room like that. A new discipline is emerging to replace them, and this series is an attempt to describe what it actually contains.

Roughly in order: how the roles have recomposed around the agent; what vibe coding costs, and the professional alternative; where decisions should live now they no longer live in humans' heads; what a mature rules layer looks like when every agent needs to read it; how review changes when the agent writes the tests; and where the discipline goes from here. Twenty essays, about thirty thousand words of prose, written in public.

The pieces will live on memex.ai. They will also land in your inbox if you subscribe below, and cross-post to LinkedIn and Medium for anyone who prefers to read there.

One piece a week for twenty weeks. The first is already written. If the case for a new discipline is one you have been waiting for someone to make, I would rather you were reading this than not.

— Barrie

The twenty essays