My AI Had Already Fixed the Code Before I Saw It

Klaassen defines compound engineering: building self-improving development systems where each iteration makes the next one faster, safer, and better. The opening anecdote is striking — before he opened his laptop, Claude Code had already reviewed a pull request, citing three months of prior feedback by PR number, without being prompted.

The distinction from typical AI engineering is temporal. Normal AI coding is prompt-code-ship-repeat with no memory. Compound engineering creates persistent learning loops where every bug fix, code review, and failed test becomes a permanent lesson the system applies automatically.

The practical example involves building a frustration detector for Cora (AI email assistant). The process follows test-driven development but adds a compound step: after each fix, the learnings get recorded so agents apply them to future work. Time-to-ship at Cora dropped from over a week to 1-3 days.

RDCO mapping: This is the clearest articulation of the knowledge compounding thesis applied to engineering. The pattern of persistent memory across sessions, self-referencing past decisions, and automated application of learned preferences is exactly what the RDCO vault and agent memory system aims to achieve for operational work.