19% slower, and sure they were 20% faster Why I wrote a book about building software with AI agents

s one point and omits the line is not an argument, it is a selection. The grades are not decoration and they are not symmetric in what they permit. A [D] claim can motivate a practice; it cannot establish one. An [A] finding on sixteen developers still generalises only as far as sixteen developers.

Taleb's Lindy effect says that for non-perishable things, life expectancy grows with age: an idea that has survived fifty years will probably outlive one that arrived last spring. Applied here it is a straightforward editorial instrument. Parnas published On the Criteria To Be Used in Decomposing Systems into Modules in 1972. Fifty-four years later it is not merely still correct, it has acquired a second and unintended application: the information-hiding discipline that was designed to limit what a human must understand turns out to be the mechanism that limits what an agent can get wrong.

That is a Lindy bet paying out. AGENTS.md is roughly eighteen months old. It may well become permanent infrastructure — Chapter 7 shows it converging bottom-up across tools and then being donated to a foundation, which is genuinely remarkable for the timescale — but you should hold the two beliefs with different grips. Read the Full Chapter This post introduces the argument.

The full chapter contains the research, evidence, and figures behind it. Read The Agentic Development Lifecycle — the complete book and Chapter 0 Next post: Same Four Letters, Opposite Advice — why ADLC can mean two different things, and why confusing them can lead you to build the wrong system. One question to leave you with: If AI could write all the code your team requested tomorrow, what would stop you from shipping all of it safely? I'd love to hear your answer in the comments.

If this is the kind of practical, evidence-driven exploration of AI engineering you want more of, subscribe to follow the series, one chapter at a time.