# Human Approval & Degrees of Autonomy

**Track:** Models, Tools & Agents — AI-Native Full-Stack Development — complete (30)
**Framework / surface:** web dev (human-in-the-loop)
**Level:** Intermediate
**Prerequisites:** Workflows vs. Agents, Identity, Sessions & Authorization

**In one line:** Autonomy is a permission gradient: read, propose, draft, modify, publish, spend, delete — each deserves a different threshold.

## Theory, aesthetics & inspiration

Human-in-the-loop is not an "Are you sure?" dialog bolted on at the end; it is the deliberate placement of approval boundaries before consequential side effects. The framing has half a century of history — Thomas Sheridan and William Verplank's 1978 taxonomy described automation as a spectrum of levels between full manual control and full autonomy, not a switch — and agentic software has rediscovered it as a permission gradient: reading public data, proposing a draft, modifying a record, publishing, spending, and deleting each deserve a different threshold. Current SDKs make the pattern concrete with tool-approval policies that pause a run, preserve its state, and resume on the human's decision — which is the important architectural idea, independent of any vendor: the decision to act must be separable from the act. Classify tools by consequence, reversibility, scope, and detectability; let the low-risk, reversible ones run free; show proposed arguments before the irreversible ones execute. And authorization still applies after approval — a user cannot approve power they do not possess.

**Builder question:** At which exact action does the cost of a mistaken autonomous decision become greater than the cost of asking for approval?