Should agentic AI autonomy expand based on earned trust?

Victor Yocco's Smashing Magazine series frames expanding an agentic system's autonomy as a staged rollout gated on demonstrated trust rather than a fixed schedule. The first phase keeps the agent to "analysis and suggestion," because "the initial goal is to build the bedrock of trust without taking significant autonomous risks." Only "once users are comfortable with the agent's proposals" does the series recommend introducing "low-risk autonomy," a phase that "is about teaching users how the agent thinks and letting them set their own pace." Full autonomy is "the final step, taken only after you have clear data from the previous phases demonstrating that users trust the system."

The gating variable at each step is evidence of trust already earned — user comfort with suggestions, then user-set pacing during low-risk autonomy — rather than elapsed time, feature completeness, or model capability. The sequence sits alongside the pre/in/post-action pattern set from the same series (see Do six UX patterns improve agentic AI user trust?): those six patterns operate within a single interaction, while this phasing operates across a product's rollout, over many interactions.

The phased model gives a temporal dimension to the role-based choice that Can we design AI autonomy separately from capability? treats as a single design decision — Yocco's series suggests a product moves a user through several of those levels in sequence rather than assigning one permanently. It also extends How should AI agents balance autonomy with user oversight? by naming a specific gating condition, "clear data" of user trust, for when decision authority should shift toward the agent.

The excerpt does not say what counts as "clear data" demonstrating trust, how long each phase should run, or what happens if trust regresses after autonomy has already been granted — the phases are described as a one-way sequence, not as a model with reversibility or failure modes built in. Whether a staged rollout actually tracks measured trust, or simply assumes it from continued use, is left open.