Why 'Write' Is Too Vague a Permission for AI Agents Acting on Code

A software developer building a GitHub adapter for AI agents found that labeling an action simply as a 'write' obscures how far its effects actually travel. Creating a pull request, merging code, sending an email, or updating production data are all technically writes, yet their operational consequences differ greatly. The author argues that flat write-access permissions can be enforced correctly while still allowing unintended high-impact outcomes. A more useful framework centers on 'reach' — how far an effect spreads before another human decision is required. Reversibility is noted as a secondary signal rather than a reliable primary boundary, since even deletable objects can leave behind logs, notifications, or consumed reviewer attention.