Collaborative Document Editing with Multiple Users and AI Agents

TL;DR

The paper addresses the challenge that current AI writing tools suit individuals, complicating team collaboration. It proposes a prototype that embeds customizable AI agents directly into collaborative documents via agent profiles and a task-driven system, with agent outputs delivered through standard comments. Through a 7-day, 30-participant study, it shows that teams treat agents as shared outputs rather than equal teammates, while agent profiles stay as personal territory; teams debate the value of one versus multiple agents and prefer manual control to autonomous action. The work contributes a UI design concept, empirical insights into team–AI interaction, and practical guidance on integrating AI as a shared resource in collaborative writing. These findings inform future interfaces that balance agent initiative with human control and reflect the social dynamics of co-writing with AI.

Abstract

Current AI writing support tools are largely designed for individuals, complicating collaboration when co-writers must leave the shared workspace to use AI and then communicate and reintegrate results. We propose integrating AI agents directly into collaborative writing environments. Our prototype makes AI use visible to all users through two new shared objects: user-defined agent profiles and tasks. Agent responses appear in the familiar comment feature. In a user study (N=30), 14 teams worked on writing projects during one week. Interaction logs and interviews show that teams incorporated agents into existing norms of authorship, control, and coordination, rather than treating them as team members. Agent profiles were viewed as personal territory, while created agents and outputs became shared resources. We discuss implications for team-based AI interaction, highlighting opportunities and boundaries for treating AI as a shared resource in collaborative work.

Figures (9)

• Figure 1: The UI for agent profiles and creation: Users create agents with descriptive text in two formats. In the structured format (A), they fill out a form-like "CV" with Name, Role, Expertise, and Skills, extendable with custom sections. In the unstructured format (B), they add free-form notes. AI suggestions for any section are available via the "Generate" button (C), and fields may be left empty. The UI also shows the list of created agents (D), including a list section with (editable) presets for quick setup. For each selected agent, the UI provides an automatically generated text summary (E) and a history of tasks (F). Clicking "Review Task" (G) reveals logs of all task runs, including the agent's selected texts, reasonings for commenting, and timestamps.
• Figure 2: The UI for task creation and activation: The task list is a fold-out sidebar (A) where users create tasks for agents by entering a title, description (instruction), assigned agent (or default "auto select"), and interaction type (manual or autonomous). For autonomous tasks (B), users choose a trigger from predefined options (elicited via a survey with academic writers, N=16). All tasks can also be run manually via the green "play" button (C). If "Create shortcut" is selected (D), the task additionally appears as a button in the floating toolbar on text selection.
• Figure 3: Example interaction flow when involving an AI agent via the comment UI: The user selects text and clicks the comment button in the floating toolbar (\ref{['fig:task_list_ui']}D) to open a new comment (A). By typing "@ai", the user brings up matching agent (or user) names (B); the user selects the default agent ("@aiauthor") and enters a request (C). An indicator shows the agent is "typing" (D) while the LLM generates a response, which then appears in the thread (E). Clicking "Append" inserts the coloured text on the page (F), which the user edits before clicking "Approve" to finalize the change and close the comment. See \ref{['fig:comments_ui2']} for further examples.
• Figure 4: Two examples of further interactions with agent responses: (A) The user clicks "..." to view the full response in a side-by-side preview, shown next to the selected text that would be replaced if accepted. This lets users read longer AI outputs in full and compare them before committing. In (B), the user asks the agent to draft the next paragraph and follows up in the same comment thread, refining the result (here by requesting a motivating example). The screenshot shows the state after clicking "Append" on the refined (2nd) response, with the new text added to the page in teal colour for review and editing before final approval.
• Figure 5: High-level overview of our findings on how teams create and interact with AI agents, tasks, and comments. From left to right: Individual users expect agents to provide value. For those focused on functional value (e.g. speed, efficiency), a single agent is sufficient (e.g. any agent can check grammar). Others value multiple agents for distinct perspectives (e.g. role play, feedback, cf. Benharrak2024aipersonas) and/or view creating specific profiles as an investment that eases future prompting. Agents and tasks are created individually but with sharing in mind: While seen as belonging to their creator, it is accepted to use them as shared resources. Users' preferred ways of using agents are those that provide control over their behaviour; either tagging an agent in a user comment on selected text (controlling where & when it responds) or triggering tasks from the task list manually (controlling when it responds).