Developer
Ask what breaks before you touch it.
The harness has its canon: guides that steer before, sensors that correct after. And three open questions — how they stay coherent as it grows, how you evaluate their coverage, what a silent sensor means. A green check that does not declare what it looked at is a harness lying by omission.
— Birgitta Böckeler — Harness engineering for coding agents · See what breaks
Ask the graph what breaks before you touch it. describe impact <node> returns the blast radius of a change — which nodes, which invariants, which ADRs — before you write a line. The codebase you just inherited can tell you what is load-bearing, and your AI agent can mount that same context instead of guessing.
The developer who lands in a codebase they did not build cannot tell what is load-bearing from what is incidental. README and docs are incomplete or stale; the originating decisions live in the heads of people who left. So the safe move is to touch as little as possible — which means feedback, fixes, and improvements stall, because the cost of an unknown blast radius falls on whoever moves first.
The same gap is what makes AI agents dangerous on an unfamiliar project: their context window cannot hold the whole codebase, so they act on a reconstructed, often hallucinated model — changing things that violate invariants nobody wrote down. Both the human and the agent are flying without a map of consequence.
The prize, once inside
When your reason for being is verifiable and not just declared, holding your course stops costing effort.