The harness around a coding agent

Since August 28, 374 commits have gone into Cold Leads. 239 of them were co-written with Claude Code.

The agent isn't the interesting part. The harness around it is:

• A memory folder in the repo: project facts, decisions and past mistakes, read before the work starts
• One rule above all: only true claims. Every statement about the product is checked against the code
• 850+ unit tests (Vitest) plus Playwright end-to-end tests as the safety net
• Nothing is pushed or deployed until I say "push"

The agent writes most of the code. I decide what gets built, read the diff and own the result.

The bottleneck moved from typing to judgment.

If you build with a coding agent: what's in your harness?