Why Linters Beat Better Prompts for AI Agents

A developer who managed three design systems at Archive.com over four years makes a sharp case: deterministic feedback loops - linters, CI, screenshot testing, runtime monitoring - matter more than model upgrades or prompt engineering when deploying coding agents at scale. The failure mode is specific. Agents produce syntactically correct code that silently violates architecture conventions, imports deprecated frameworks, uses magic numbers instead of design tokens, and adds console.log where Datadog loggers belong. The numbers back it up - CodeRabbit found 2.74x more security vulnerabilities in AI-generated code, while Spotify's Honk agent merges 650+ PRs monthly only after three years of feedback infrastructure investment. The progression from manual review to self-tightening loops where agents, custom lint rules, CI, and observability feed into each other is becoming the practical playbook for teams serious about shipping agent-written code at scale.