Writing on AI Governance, Agentic Safety Architecture, Design System Migrations, and Performance Engineering.
My safety doc said a control was enforced at the tool layer. It was enforced by nothing, and a passing test in the same repository had contradicted the doc for 21 days. The documented-but-unenforced control as a defect class.
A reviewer that knows who wrote the code and why reviews it worse. Independence comes from withholding provenance, not adding context. Your AI reviewer should know less.
We scaled code generation and left review behind. Four weeks of org data: 90.7% of pull requests AI-assisted, and at least half merged with no human review. The dashboards say review caught up because they count bots as reviewers. The gap, and how to close it.
Automating a bug pipeline is the easy part. Deciding which tier each bug belongs to is the bottleneck, and it is where the real engineering lives.
A controlled ablation, 42 runs: a behavioral-payload skill added zero marginal defect detection over a no-guidance control. A skill whose payload is documentation inherits documentation's 0%.
Unrestricted shell access handed to an AI agent, then the defense-in-depth architecture that made it survivable: layered hooks, and the self-approval bypass patterns they had to close.
T7,798 generated components in one package. The real cost of the barrel file is not bundle size (it tree-shakes fine), it is build-time parse, plus a dual type-resolution trap that runs but will not type-check.
Two design systems on one page usually means a rewrite. A two-tier CSS architecture, scoped plus utilities, lets them coexist without one.
A migration guide starts rotting the moment it is written. Separate knowledge, procedure, and data, and the guide becomes a product that stays right.
A controlled ablation across nine tasks: a fully-specified CLAUDE.md changed agent behavior 0% of the time, identical to having no rules at all. The data, and why documentation cannot govern behavior.
