Pattern: Willison's Trust Gradient — Treat Agent Output Like Another Team's Code, Not Your Own
Simon Willison·medium signal
Willison's May 6 blog articulates a trust model emerging among experienced practitioners: stop reviewing every line of agent-generated code. Instead, treat it like code from another team in a large org — a semi-black box you trust until it breaks. The key shift: evaluate agent output by proof-of-use (did it actually work in production?) rather than code-quality metrics (test coverage, documentation). This lowers the review burden enough to capture the 10x productivity gain without pretending you're auditing every token.