What action patterns emerge when humans and agents edit concurrently?

A Study 2 design probe, Cleo, "tracks user actions and selectively updates its plan" in a shared Figma canvas, built after an earlier probe found that unacknowledged user edits discouraged intervention. With 12 professional designers, the authors "analyzed 255 interaction turns and stimulated recall interviews, identifying five action categories, ten codes, six triggers, and four enabling factors." Concurrent interaction — the user acting while the agent was still executing — occurred in "33.73% of turns." The five categories include direct edits, where a user edits an element to demonstrate a change for the agent to propagate or make a purely local change; "Artifact takeover, in which users duplicate work-in-progress"; "Opportunistic takeover," where "the user completes a pending subtask" the agent had not yet reached; and "Intermediate result appropriation," where the user reuses "partial outputs" the agent produced mid-task.

Each pattern, in the authors' account, implies a different response the agent should make rather than a single uniform rule. A direct edit requires the agent to judge "intended scope... before carrying changes into subsequent operations" — whether to treat it as local or to propagate it. An artifact takeover should prompt the agent to "preserve separate work areas so that the user can refine the copy while the agent continues the original." An opportunistic takeover means the agent "should account for that contribution in its remaining plan" rather than redoing finished work. Intermediate-result appropriation means the agent should "preserve those outputs while continuing its assigned work" instead of overwriting what the user has already taken. The paper frames this as the agent needing to read "which artifact or subtask the user acts on and what changes" to tell correction, parallel work, task-sharing, and reuse apart.

This extends Can users steer AI agents better by revising plans mid-execution? into execution-time, unplanned territory: Cocoa structures delegation as an explicit, assignable choice per plan step before and during a document task, while this taxonomy describes division-of-labor patterns that designers generated spontaneously, mid-execution, in a visual canvas, without any prior assignment step. It also answers the gap named in Does visible agent progress help designers know when to intervene? — that probe's designers wanted to intervene but were discouraged when the agent misread their edits; this taxonomy is the vocabulary Cleo uses to tell those edits apart and respond to each appropriately. It is also an instance of the broader point in How do communication modalities shape human-agent collaboration patterns?: varying voice interaction, process transparency, and presence awareness as deliberate controls is what let the authors observe these distinct behavioral patterns rather than an undifferentiated stream of "intervention."

The taxonomy comes from one probe system — a ReAct-based Figma agent using Claude Sonnet 4.5 and Claude Haiku 4.5 — tested with 12 participants over two days; the authors state this "requir[es] larger-scale longitudinal validation" and flag that different architectures (e.g., reflection-based, hierarchical planning, continuous execution) may elicit different interaction patterns. The 33.73% figure and the five categories describe this study's sample of turns, not a general rate or an exhaustive list of how concurrent human-agent interaction unfolds across tools; the implication the authors draw is that agents sharing a workspace need some version of this scope-reading capability, not that these five categories are the final set.