A detailed account of how one team runs full agentic development in production, where no application code is written by hand. The methodology centers on three coordinated layers: non-negotiable root rules (AGENTS.md/CLAUDE.md), area-scoped rules loaded contextually, and a workflow skill chain (brainstorm → plan → subagent execution → review → finish). Testing is enforced as defense-in-depth with four mandatory layers: static CI checks, unit/integration tests with Vitest, real-stack Playwright API tests, and UI smoke/regression tests. TDD is strictly enforced — code written before its test is deleted. Documentation is treated as a first-class output with freshness checks. The post also covers the mindset shift required, the growing importance of the conceptual/architecture phase, and how the methodology scales to higher-stakes regulated domains with additional gates like mutation testing, immutable audit logs, and dual-model parallel validation.
Table of contents
Why convention is not a style preference anymoreThe three layers that keep increments coherentDefense in depth: our testing strategyTDD with agents: why it matters more, not lessDocumentation that stays aliveThe mindset shift nobody tells you aboutWhen the stakes are highWhat we've learned, so farSort: