Amazon’s 2015 shareholder letter splits decisions by reversibility. Type 1 / one-way door: consequential and irreversible or nearly so, and it “must be made methodically, carefully, slowly, with great deliberation and consultation.” Type 2 / two-way door: changeable, reversible, “you can reopen the door and go back through” - and it should use a light-weight process, decided by high-judgment individuals or small groups.
The failure mode the letter names is applying the Type 1 process to Type 2 decisions, which produces “slowness, unthoughtful risk aversion, failure to experiment sufficiently, and consequently diminished invention.”
The agent corollary
An agent makes execution nearly free while leaving the cost of a wrong call exactly where it was. So reversibility stops being an organizational nicety and becomes the routing rule for what you delegate.
| Door | Route |
|---|---|
| Reversible, contained blast radius | Delegate now, no deliberation |
| Reversible, wide blast radius | Delegate with the reversal path written first |
| Near-irreversible | Premortem, then a human runs it |
Two adjustments the original framing does not cover:
- Reversibility is manufacturable. A branch, a feature flag, a down migration, a dry run, or a build-time check usually converts a one-way door into a two-way one for less than the deliberation would have cost.
- Reversibility decays with repetition. Two hundred individually reversible decisions, taken independently and never revisited, become an architecture nobody chose. At agent speed this accumulates faster than anyone reviews it.
Why it goes in the harness
A door classification the model is merely told about is not a classification - it is prose the model re-weighs every turn. Express it as a permission rule or a hook instead: deny rules are evaluated before the permission mode, so they hold even under bypassPermissions.
Full treatment, with the six routing rules and a paste-in gate: One-way doors your agent can already walk through.