Approval gate

Added · Seed

Seeded from a cross-catalogue triage of six AI pattern libraries (references/ai-pattern-catalogues.md); attested in all of them, but not yet written from close product study.

Machine-initiated work pauses at a gate until the actor consents — per action, per batch, or per diff — so autonomy never outruns the actor’s willingness to answer for the result.

Forms

Recurring forms: approve/reject on a single proposed action, batch review of queued actions, diff review of proposed changes, and graduated modes (suggest, confirm, execute) that move the gate rather than remove it.

Resources & references

To-do

The rubber-stamping force deserves evidence (overlaps the overreliance literature already cited on cognitive forcing functions). Where batching turns review into problem-curated view territory is worth mapping.

Consequences

  • nothing consequential happens without a human yes, which buys safety at the price of throughput: the agent now works at the speed of the actor's attention
  • each gate is an interruption, and past some frequency approving stops being a judgement and becomes the habit the gate was meant to prevent
  • the approval concentrates accountability: whoever clicked yes now owns an outcome they may have reviewed only glancingly

Related patterns

Serves

  • Delegation — the execution-boundary checkpoint of delegation's lifecycle

Enacts

  • Agency — keeps initiation and consent in different hands

Complements

Alternatives

  • Undo — consent before the act and recovery after it are the two ends of one safeguard ladder

Related

  • Inline confirmation — the single-action rung of the same ladder; an approval gate reviews delegated work, often in batches
  • Action consequences — supplies the severity reasoning for where a gate belongs
  • Suggestion — accepting proposed content versus authorising an act — the same accept/reject core at different stakes

Preceded by

  • Bounded autonomy — crossing a bound is a designed event rather than a failure: the work pauses and control returns to the actor for renewal, escalation, or takeback
  • Change review — the delta is a reviewable object — what was, what would be, side by side or in place — so judgement lands on the change rather than on re-reading the whole, and an approval has something concrete to act on
  • Plan preview — the agent's intention is now a reviewable object — steps, targets, expected effects — so misreadings surface while they are still cheap, and the preview becomes the thing an approval can act on