Do users develop stable collaboration styles with web agents?

A corpus named COWCORPUS, built from "400 real-user web navigation trajectories containing over 4,200 interleaved human and agent actions" (2,748 agent action steps and 1,476 human action steps), finds that when real users work alongside a web-browsing agent, their behavior does not vary unpredictably task by task but "combine[s] into consistent higher-level collaboration strategies." The paper groups users into four styles — "Takeover, Hands-on, Hands-off, and Collaborative" — that "capture how users balance supervision, intervention, and share control with the agent." Tasks mixed "10 standard tasks from the Mind2Web dataset" with "10 free-form tasks of the participants' own choice," so the styles were observed across both benchmark and self-chosen tasks rather than one task type.

The paper traces the styles back to three recurring motivations behind individual intervention events — "error correction, preference refinement, and assistive takeover" — and argues that these motivations, repeated across a session, harden into a stable collaboration style rather than remaining one-off reactions. Step-level annotations mark "when users intervened by pausing, resuming, or overriding agent execution," which is how the four styles were derived empirically from the trajectory data rather than assumed in advance.

This taxonomy describes variation in the people being designed for, which sits upstream of two adjacent design questions already in the vault: Can users steer AI agents better by revising plans mid-execution? gives one concrete mechanism — per-step delegation — that a Collaborative-style user could lean on, while Does oversight strategy prevent risky agent actions or just catch them? compares interface strategies that structure intervention for users in general, without distinguishing style. Both studies hold the interface fixed and vary structure or risk; this one holds the task fixed and varies the person, suggesting style is a missing variable in both. It also extends the people-side concern in What UX principles do workplace users want in AI agents?, which names control as one UX principle generally wanted, by showing wanting control is not uniform but splits into at least four distinct patterns.

The excerpt does not establish whether these four styles are stable traits of a given user across different tasks and over time, or situational responses to a particular task's stakes and familiarity — the dataset captures style within a session, not whether the same person adopts a different style on a different task. It also does not report what share of the 400 trajectories fell into each of the four styles. The implication, at the strength the evidence supports, is that an agent's intervention policy may need to be conditioned on which style a given user or session exhibits, rather than built around a single generic intervention threshold applied to everyone.