Agent usability tests need recovery tasks, not completion ratings

Agent usability tests need recovery tasks, not completion ratings. I learn little from asking whether an agent eventually produced an answer or whether a participant liked the experience. Production trust is formed after the first wrong turn—when a person must understand what happened, decide whether to intervene, repair the work, and resume without starting from scratch.

Traditional completion measures flatten that sequence into one bit. An agent may finish while silently violating the user’s intent, duplicating an effect, or forcing the user to reconstruct hidden state. Another may stop safely, show the failed step, and make correction cheap. A completion score calls the first system successful and the second unsuccessful even though the second is the one I would trust with consequential work.

Usability simulation can make this richer before a human study begins. Persona-driven LLM agents can interact with a real web design at scale, using simplified HTML observations and a task-independent action vocabulary. A framework can vary demographic profiles, preserve a memory stream weighted by importance, relevance, and recency, and replay each session step by step. That is useful for finding where different simulated users hesitate or diverge.

But the test script must inject failure deliberately. I would expire a credential halfway through a task, withhold a required field, return conflicting tool evidence, interrupt a long operation, or make the proposed action exceed its approval scope. Then I would ask whether the participant can identify the blocked state, locate the consequential decision, revise the input, and resume from a known checkpoint. Recovery is not an edge case added after the happy path—it is the interaction model under realistic conditions.

The replay interface matters because aggregate ratings hide causality. A step-through trace can show the observation available to the participant, the action selected, the remembered context, and the point at which the interface ceased to support a sound decision. An interview with a memory-loaded simulated user can probe why the person believed the task was safe or complete. That evidence connects interface choices to behaviour instead of treating “confusing” as an explanation.

I would score recovery as a sequence of observable states: failure recognized, blast radius understood, authority correctly transferred, repair option selected, duplicate effects avoided, and useful progress preserved. Confidence should matter only when it changes the handoff. A low-confidence warning that leaves the same actor, permissions, and next action in place is decoration, not oversight.

This also changes what designers expose. Reasoning text is optional; recovery state is not. The interface must name what succeeded, what remains uncertain, what was not attempted, and what the user can safely do next. That is why recoverability beats explainability: a persuasive account of failure does not itself restore control.

There is one precise concession: synthetic participants can cheaply widen scenario and persona coverage, but they cannot establish human usability or represent lived stakes. Their behaviour depends on the model, persona generator, memory design, and connector. I would use them to discover candidate failures and improve protocols before human studies—not to replace observed human evidence or claim demographic validity.

The acceptance question is therefore not “did most agents finish?” It is “when the system became uncertain or wrong, could a person recover intent and control without creating a second failure?” A usability programme that never asks participants to repair the agent is measuring a demo. Production agent UX begins where completion stops being guaranteed.