Does visible agent progress help designers know when to intervene?

A Study 1 design probe with 10 professional UI/UX designers tested a Figma plugin agent that exposes "its design process in real-time through step-by-step execution, so users observe not just what the agent produces but how it progresses." The authors report that "participants used visible progress to evaluate the agent's direction and identify opportunities to contribute," and that they "wanted to reuse intermediate results and demonstrate intent through direct manipulation." But the same visibility cut the other way: "when users edited the shared canvas while the agent was executing, the agent sometimes misinterpreted or reverted their modifications, creating coordination conflicts and discouraging further intervention." The authors note this "reluctance to intervene limited us to observe the collaboration strategies participants wanted to pursue" — the conflict suppressed the very behavior the probe was built to study.

The probe grounded its design in three components the authors say are "drawn from human-human collaboration literature": Interaction Modality (voice input with direct canvas-object reference, for "natural turn-taking" and "a sense of horizontal partnership"), Process Transparency (the step-by-step visible execution), and Presence Awareness (a floating agent character showing focus, following "Gutwin's workspace awareness framework"). The agent itself runs a ReAct reason-act-feedback cycle over 38 Figma operations, narrating each step and self-evaluating plan completion. Process Transparency and Presence Awareness gave designers the signal to find openings; what the agent lacked was a way to take their edits back in — the paper's gap diagnosis is "the need for collaborative context awareness on the agent's side, so it could respond to users' contributions in the ways they expected."

This cuts against How can users verify what an AI agent actually plans to do? from the other direction: that paper frames common ground as the user needing a channel to verify the agent's plan; this probe shows a symmetric failure where the agent could not ground on the user's edit, breaking the shared understanding from the agent's side instead. It also sits next to Do proactive AI coding agents actually disrupt programmer workflow? — that study found presence indicators ease disruption from a proactive agent; here, presence and progress visibility were present yet still produced conflict, because the missing piece was the agent's response to the edit rather than the user's awareness of the agent. Can users steer AI agents better by revising plans mid-execution? addresses a related problem — letting users and an agent interleave work — by making delegation an explicit, assignable per-step choice in a document; this probe's canvas lacked that explicit hand-off mechanism, which is why edits landed as unexpected interruptions rather than assigned work.

The study is a single-session, exploratory probe with 10 participants and does not quantify how often misinterpretation occurred or isolate which agent behaviors (reverting vs. simply ignoring an edit) caused the reluctance. It motivated a second probe, Cleo, built specifically to track and incorporate user actions — see What action patterns emerge when humans and agents edit concurrently? — rather than supporting a general claim that process transparency backfires; the implication the authors draw is narrower: transparency creates openings for contribution only if paired with a mechanism for the agent to incorporate what users do with those openings.