
Enforcing AI agent rules via git worktree isolation and mechanism-layer hooks
Original: The day I moved the rules to a place they can't be broken even if you try
Short summary
When multiple AI agents work on the same repository, document-based rules get ignored as sessions grow long, causing accidents like lost changes and file collisions. The author solved this by moving rules into a mechanism layer—git pre-execution hooks, hosting-side branch protections, and worktree-based isolation—that enforces compliance without relying on agent judgment. Across three repository groups the pattern converged: detached-HEAD worktrees for read-only investigation and branch worktrees for PR-producing work, each created in 0.25 seconds.
- •Document-based rules for AI agents fail as sessions grow long; mechanism-layer enforcement (hooks, hosting settings, worktrees) prevents violations structurally
- •Git worktree isolates each agent into an independent working directory in 0.25s, sharing commit history but keeping local state separate
- •Pattern converged across three repo groups: detached-HEAD worktrees for investigation, branch worktrees for PR work
Generated with AI, which can make mistakes.
Is this a good recommendation for you?



