Agent Harness
The root Claude Code setup every other project inherits: the shared constitution, the skills, and the build loop that runs a spec to done.
CREATED JUN 15, 2026 · UPDATED AUG 14, 2026
PROBLEM
Every repo was re-teaching the same agent the same things: the house voice, the design tokens, when to escalate instead of guessing, what a finished change looks like. The instructions drifted apart repo by repo and the worst version was always the one being used.
HYPOTHESIS
One root configuration, inherited rather than copied, keeps every project's agent working to the same standard — and makes improving that standard a one-file change.
ARCHITECTURE
A global constitution plus per-domain skills, layered under each repo's own AGENTS.md. Shipped alongside is a build loop that takes a written spec through scout, plan, build and audit, refusing to start a milestone until the previous audit passes.
STACK
RESULT
In use by every repository in this lab. aeroslip., collect. and this site all run on it.
LEARNINGS
- Agent instructions are code and rot like code — one copy, inherited, or they diverge.
- The rule that pays for itself is the one about when to stop and ask.