Bounded autonomy

Added · Seed

Seeded from a cross-catalogue triage of six AI pattern libraries (references/ai-pattern-catalogues.md); the agentic wave's most attested new move, not yet written from close product study.

The actor grants the system authority with explicit edges — how much it may spend, touch, or decide, and for how long — so delegation can start small and grow with demonstrated competence.

Forms

Recurring forms: autonomy levels per task (suggest / confirm / execute), budgets in time, actions, or money, per-capability rather than per-application grants, and thresholds that automatically demote autonomy when risk rises.

Resources & references

To-do

The relationship between bounds and earned trust (widening limits over a history of good runs) touches ai-tuning and delegation’s trust arc; map it when the delegation foundation’s trust prose lands.

Consequences

  • granted authority now has explicit edges — scope, spend, time, risk class — so trust can be extended in increments instead of all at once
  • 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
  • the bounds only bind if the actor understood them at grant time; authority granted through an unread scope screen is consent in name only

Related patterns

Precedes

  • Handoff — 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
  • Approval gate — 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

Serves

  • Delegation — bounded autonomy covers delegation's set-up station — the shape of the authority granted before the agent runs

Enacts

  • Agency — autonomy becomes something the actor meters out rather than switches on

Complements

Related

  • Settings — where standing bounds tend to live between grants