AXD for Designers

Design delegation confirmation dialogs that clearly state the scope of agent authority - what the agent can and cannot do.

Use progressive disclosure to reveal agent capabilities gradually, building trust before expanding autonomy.

Implement visual scope indicators that show users the current boundary of agent authority at all times.

Design explicit 'hand back' controls that allow users to reclaim authority from an agent at any point in a workflow.

Use distinct visual language for agent-initiated actions versus human-initiated actions to maintain legibility.

Design confidence indicators that communicate agent certainty levels without overwhelming users with probabilistic data.

Implement action preview patterns - showing users what an agent is about to do before it acts, with a confirmation step for high-stakes actions.

Use audit trail visualisations that make agent action history legible and reviewable at the appropriate level of detail.

Design trust calibration flows that gradually expand agent autonomy as users demonstrate comfort with agent behaviour.

Avoid designing agents that appear more confident than they are - false certainty is a primary trust failure mode.

Design explicit 'agent unavailable' states that clearly communicate what has happened and what the user can do next.

Implement human fallback paths for every agent-mediated workflow - there must always be a way to complete the task without the agent.

Use status communication patterns that keep users informed of agent availability in real time.

Design recovery flows that allow users to resume interrupted agent workflows from a known-good state.

Test absent-state designs with users who have never seen them before - absent states are often under-designed because they are rarely encountered in testing.

Design outcome definition forms that capture success criteria, acceptable ranges, and explicit failure conditions.

Use natural language input with structured confirmation - allow users to describe outcomes conversationally, then confirm in structured form.

Implement outcome preview patterns that show users what a successful outcome looks like before the agent begins.

Design constraint specification interfaces that allow users to define what the agent must not do, not just what it should do.

Use iterative refinement patterns - allow users to adjust outcome specifications mid-task without restarting the workflow.