← Principles
validated

“Prefer reversibility to permission.”

Every permission gate spends human attention; reversibility spends none until something goes wrong. Where actions can be made cheap to undo — checkpoints, sandboxes, rewind — autonomy can be granted freely and review moves to accept-time. Gates should be reserved for what cannot be taken back.

The Context Matrix

Where this principle has been tested. Untested cells are honest gaps, not passing grades — seen evidence for one? Tell us.

ChatIDECLICanvas / DocAmbient / Background
Coding ● ●
Knowledge Work ●
Consumer ◐
Enterprise Ops
Creative

● supports · ◐ boundary · ✕ violates · blank untested

Linked Patterns

RecoveryApprovalMemory

Boundary Conditions

  • consumer × chat × auto-with-gates

    Browser agents acting on the live web (purchases, messages, form submissions) cannot be sandboxed into reversibility — permission gates remain the right tool where actions are irreversible by nature.

Evidence

supports

Claude Code checkpoints + rewind make edits cheap to reverse, so permission modes can loosen.

coding × CLI × supervised · Claude Code: checkpoints and rewind

supports

Codex sandboxes entire runs: full autonomy inside the sandbox because discarding the diff is free.

coding × ambient-background × background · Codex: sandboxed runs as implicit recovery

boundary

Browser agents acting on the live web (purchases, messages, form submissions) cannot be sandboxed into reversibility — permission gates remain the right tool where actions are irreversible by nature.

consumer × chat × auto-with-gates · ChatGPT agent mode: take over and stop controls

supports

File-based memory is inherently reversible (git, undo) — no permission ceremony needed to let the agent write it.

knowledge-work × IDE × supervised · Claude Code: CLAUDE.md and auto memory