The Compound Effect of Guardrails in the Era of AI

In this talk, I walked through a development workflow where a request moves from Slack to Linear, an AI agent reads the codebase and opens a pull request, and automated reviews and CI run before engineers make the final call. The speed of that loop makes the quality of the codebase matter more: every pattern an agent finds becomes an example for the next change, so both good decisions and weak foundations accumulate over time.

I broke down seven guardrails that shape those decisions: instructions, context, types, lint rules, formatting, runtime schemas, and tests. I spent particular time on making lint fast enough to keep enabled, introducing Biome rules gradually, and enforcing conventions in CI. I also covered setting a test coverage floor and raising it over time. The central idea is that each small rule keeps working across future changes, giving teams a more reliable foundation for AI-assisted development.