← Autonomy Foundry

Human approval should be triggered by consequence, not by uncertainty

An agent should not ask a human about every ambiguous detail. Approvals are most useful when an action crosses a consequence boundary: money, legal commitment, irreversible publishing, sensitive data or privilege escalation.

Good approval triggers

Use explicit thresholds for spend, production changes, external communications, permission grants, deletion, contract acceptance and access to sensitive systems.

Keep the queue small

If everything requires approval, autonomy disappears. If nothing does, accountability disappears. Define a narrow set of high-impact actions and let low-risk reversible work proceed automatically.

Record the decision

Every approval should preserve who requested it, what action was proposed, expected impact, cost, evidence and who approved or denied it.

Run the free autonomy risk check →