cs.HCSep 24, 2026
SaveDocuTeam: Mixed-Initiative Multi-Agent Discussions around Evolving Documents
Organizations: School of Computing, KAIST, Daejeon, Republic of Korea · College of Liberal Studies, Seoul National University, Seoul, Republic of Korea · SkillBench, Santa Barbara, CA, USA
Abstract
In open-ended problem solving, collaborators often rely on discussion to surface concerns, challenge perspectives, and refine shared work as it evolves. While AI agents are increasingly used as discussion partners, existing multi-agent systems place a heavy burden on users to initiate and carefully orchestrate the discussions. We present DocuTeam, a mixed-initiative multi-agent discussion system in which both users and agents can initiate and steer conversations. Agents monitor document changes to proactively start and redirect discussions as the work evolves, while users can flexibly shape the conversation or adopt agent ideas. In a within-subjects study (N=20), participants using DocuTeam produced outcomes rated significantly more novel, relevant, and specific than with a baseline without any increase in cognitive load. Rather than using agents for one-off idea sourcing, participants engaged in an iterative refinement loop in which document changes prompted agent reactions, which led users to revisit and further develop their work.
Figures & tables
Figure 1 . Illustration of a mixed-initiative multi-agent discussion interface. Whereas human teammates can organically build on and challenge one another’s perspectives around shared work, existing multi-agent discussion systems often require users to initiate and orchestrate discussions separately from their working context. DocuTeam supports (A) organic agent-initiated discussions that proactively surface issues, (B) mixed-initiative discussion steering that allows both users and agents to guide the flow of conversation, (C) discussion sessions anchored to relevant document regions, and (D) selective incorporation of ideas from the discussion. The figure contrasts human collaboration and existing multi-agent discussion systems on the left with DocuTeam on the right. At the top left, several human collaborators discuss an evolving shared document, illustrating discussion that emerges around ongoing work. At the bottom left, a conventional multi-agent interface separates the document from a side-panel chat and requires the user to initiate and sustain discussion. In DocuTeam (right), agent discussions are anchored directly to specific regions of an AI Workshop Plan document, demonstrating four core interactions. Agents are dragged onto a target document block and autonomously discuss its content beside that region, here debating whether to drop the big-name speaker and redirect the budget to Activity 2 (C). A discussion snippet, "Mini-hackathon with prizes," is applied into the TBD schedule slot and adapted to fit the surrounding context (D). Agents proactively initiate a discussion on an unresolved group formation policy line marked with question marks, surfacing questions like "Random group?" and "How many attendees?" without user prompting (A). When agents need user input to proceed, they surface a clarification question directly in the document, here asking how many people will join the event with selectable options (30–50, 50+) and a text input (B).
| Collaboration Systems | Agent-Initiated Interaction | Autonomous Discussion Development | Grounding in Evolving Work |
| Single Agent ( Son et al., 2025 ; Pu et al., 2025b ; Prasongpongchai et al., 2025 ) | Yes: An agent proactively offers assistance and collaborates with the user. | No: No agent-to-agent discussion. | Yes: An agent accesses the document and observes users’ behavior. |
| Set of Multiple Agents ( Choi et al., 2025 ; Chen et al., 2026 ; Lehmann et al., 2026 ; Fang et al., 2025 ) | Limited: Agents are invoked by user requests or predefined conditions. | Limited: Agents may exchange or synthesize perspectives, but do not autonomously initiate and develop open-ended discussions. | Yes: Each agent’s contributions are grounded in the work artifact or context. |
| Multi-Agent Discussion ( Quan et al., 2025 ; Liu et al., 2025b ) | Limited: Users initiate the discussion and manage when it occurs. | Partial: Agents debate, critique, and build on one another’s perspectives, but require users’ orchestration (e.g., via @mention or a “Continue” button). | Limited: Discussions are detached from the evolving work artifact. |
| DocuTeam | Yes: Agents proactively initiate discussions without requiring explicit user queries. | Yes: Agents autonomously develop and direct discussions with one another. | Yes: Discussions are grounded in the shared document as it evolves. |
Table 1. Comparison of interactions in Human-AI collaboration systems for document work. Unlike prior systems, DocuTeam enables agents to proactively initiate and autonomously develop multi-agent discussions grounded in evolving work. Table comparing four types of human-AI collaboration systems—single agents, sets of multiple agents, existing multi-agent discussion systems, and DocuTeam, across agent-initiated interaction, autonomous discussion development, and grounding in evolving work. Recent single-agent systems are proactive and grounded but lack agent-to-agent discussion; sets of multiple agents offer limited initiative and discussion development; existing multi-agent discussion systems support agent-to-agent discussion but typically require user initiation and are weakly grounded in evolving work. DocuTeam supports all three capabilities.
Figure 2 . DocuTeam overview. DocuTeam consists of two main components: (A) a Document overlay view, where discussion sessions are anchored to specific regions of the document, and (B) a Side Panel, where users manage agents and inspect ongoing discussions. Users can initiate discussions by dragging agents onto the document and selecting a discussion mode (E, F). Discussions appear as anchored discussion bubbles beside the relevant document region (G) and can be opened in the side panel as chat threads with summaries (C, I). Agents proactively attend to edited blocks (H) and start discussions, while users can apply useful discussion content back into the document (J). System overview of DocuTeam showing two main components. The left side displays an AI Workshop Plan document with discussion bubbles anchored beside specific blocks, showing agent avatars and message previews; one bubble has a red border prompting the user to choose a focus. The right side panel lets users manage agents, read ongoing chat threads, and view structured summaries of discussions. Users can start a discussion by dragging agents onto a document block and selecting a mode. A floating agent avatar beside an edited block indicates proactive monitoring. An Apply dialog lets users insert a selected discussion snippet directly back into the document.
Figure 3 . Blind evaluations showed that the DocuTeam condition produced more novel, relevant, and specific outcomes than the Baseline condition. The mode of user-initiated discussions shifted from Idea to Discussion. (*:p<.05, **:p<.01, error bars indicate one standard deviation) Two grouped bar charts comparing Embedded (blue) and Detached (gray) conditions, with error bars indicating one standard deviation. The left chart shows blind evaluation scores (0–5 scale) across four quality dimensions. Embedded scores higher than Detached on Novelty (3.20 vs. 2.78, p<.05), Relevance (3.30 vs. 2.78, p<.05), and Specificity (3.25 vs. 2.28, p<.01), while Workability shows no significant difference (3.4 vs. 3.28). The right chart shows the frequency of user-initiated discussion modes (0–5 scale). Embedded participants used the Idea mode less than Detached (1.53 vs. 2.8, p<.05), used Discussion mode more (2.1 vs. 1.2, p<.05), and Evaluation mode was similarly low in both conditions (approximately 0.6).
Figure 4 . Post-survey responses on outcome and collaboration. Unlike the external evaluators’ assessments, participants’ confidence in and satisfaction with their own outcomes did not differ significantly between conditions. DocuTeam made discussion ideas easier to apply to the document. (*:p<.05) Diverging stacked bar charts using a 7-point Likert scale from Strongly Disagree (dark navy) to Strongly Agree (dark red), comparing DocuTeam and baseline conditions. It shows six outcome and collaboration items. Outcome Confidence and Outcome Satisfaction show similar distributions across conditions, both skewing toward agreement. Felt like a tool than collaborator (reversed) are similarly distributed between conditions. Easy to apply chat ideas is the only significant item (p<.05), with DocuTeam scoring higher than baseline, indicating DocuTeam made it easier to apply discussion ideas to the document.
| Metric | DocuTeam | Baseline |
| Steering msgs.* | ||
| Draft word changes.** | ||
| Msgs applied to doc.** |
Table 2 . Differences in task process measures between Baseline and DocuTeam . Table comparing three task process metrics between the treatment and control conditions. The treatment group sent fewer steering messages than the control group (2.70 vs. 5.00), suggesting less need for correction. However, the treatment group made more word-level draft changes (448.70 vs. 322.95) and applied more messages to the document (12.30 vs. 9.00), indicating more active document editing. All values are means with standard deviations. Differences in steering messages are significant at p < .05; draft word changes and messages applied to document are significant at p < .01.
Figure 5 . NASA-TLX workload ratings. Overall perceived workload and performance remained similar. Diverging stacked bar charts using a 7-point Likert scale from Very Low (dark navy) to Very High (dark red), comparing DocuTeam and baseline conditions. It shows five NASA-TLX workload dimensions. Mental Demand, Temporal Demand, Effort, Performance, and Frustration all show broadly similar distributions between DocuTeam and baseline, with no significant differences, suggesting overall workload was comparable across conditions.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6 . Detailed expectations for multi-agent support across task stages in the formative study. Rows indicate stages of work, and columns indicate three types of support participants expected from multi-agent discussion: resolving conflicting objectives, validating emerging outcomes, and synthesizing new concepts. Each cell summarizes a representative expectation with participant identifiers. Table mapping four design process stages (rows) against three multi-agent collaboration strategies (columns). Rows are: Set goal/criteria, Exploration and Formulation, Collection (Solution), and Presentation (Final review). Columns are: Resolving conflicting objectives, Validate the outcome, and Synthesize new concept. In the Set goal/criteria row, agents discuss and prioritize factors (G1P1, G3P3) to resolve conflicts; clarify ambiguous criteria from multiple aspects (G1P1) and quality-check user-written criteria (G4P3) to validate outcomes; and suggest diverse or unconsidered goals (G1P3, G2P3) to synthesize new concepts. In the Exploration and Formulation row, agents select and debate which information is important (G3P2); mutually fact-check information (G1P2); and provide insights from diverse perspectives and explore underexamined areas (G1P2, G3P1). In the Collection (Solution) row, agents representing different solutions debate each other and weigh the most reasonable option given the user's circumstances (G2P3); assign ratings or scores to each solution condition (G5P3); and each propose different design decisions or alternatives (G2P1). In the Presentation (Final review) row, each agent represents a specific evaluation criterion and comments accordingly (G4P2); personas such as clients, designers, or developers identify issues in the final design (G1P1); and agents act as readers at different expertise levels and adapt their explanations accordingly (G2P2).
Figure 7 . Summary interface for each discussion mode. Three-panel interface showing outputs of Idea, Discussion, and Evaluation modes. The Idea panel displays a summary header stating the conversation is about brainstorming a keynote concept within budget, followed by bullet-point ideas grouped into three categories: Speakers Union (e.g., duo talk between AI scientist and rising researcher), Form and Impact (e.g., pitching and mentoring format), and Reduce Costs and Leverage Your Budget (e.g., performance-based compensation contract). The Discussion panel shows a summary header about organizing small round tables, then lists three consensus items (e.g., labeling tables by field or experience level) and two conflict items with sub-points representing opposing views, such as whether to group similar or mixed disciplines at roundtables. The Evaluation panel shows a summary header about three agents providing feedback on labor costs and facilitator issues, followed by structured feedback from three named agents: Budget Keeper (flagging unaccounted facilitator and recorder costs), Organizer (warning that missing facilitators reduce discussion quality), and AI Professional (recommending standardized output formats and one-page summary templates). Each panel includes a colored vertical accent bar and expandable arrow icons on each item.
Figure 8 . Participants’ responses to the remaining six survey items across the DocuTeam and Baseline conditions. Scores were measured on a 7-point Likert scale (1: Strongly Disagree, 7: Strongly Agree). (**:p<.01) Diverging stacked bar chart showing responses to six survey items comparing DocuTeam and baseline conditions on a 7-point Likert scale, where 1 is Strongly Disagree and 7 is Strongly Agree. Colors progress from dark navy (1) through light blue, gray, orange, burnt orange, to dark red (7). For "Expanded thinking," both conditions skew toward agreement, with DocuTeam responses more concentrated at the higher end of the scale (5: 5, 6: 10, 7: 4) and baseline responses more spread across the scale (1: 2, 4: 3, 5: 5, 6: 7, 7: 2). For "Deepen thinking," responses were mixed in both conditions, although DocuTeam showed somewhat more responses at the higher end (5: 4, 6: 7, 7: 3), while baseline included more responses at the lower end. For "Willing to use this again," both conditions skew toward agreement, with DocuTeam responses concentrated in the middle-to-high range (5: 3, 6: 11, 7: 4) and baseline slightly more spread (3: 2, 4: 2, 5: 4, 6: 7, 7: 3). For "Difficult to be mentally engaged," both conditions lean toward disagreement, with DocuTeam showing 4 and 7 at the low end and baseline showing 3 and 7, suggesting participants did not find the tasks mentally disengaging. For "AI was aware of my actions" (significant at p<.01), DocuTeam responses are more spread across mid-to-high scores (2: 2, 4: 4, 5: 5, 3: 3, 5: 5) while baseline leans higher (3: 3, 6: 6, 5: 5, 4: 4), indicating baseline participants felt the AI was more aware of their actions. For "I was aware of AI's actions," distributions are similar across conditions, with both showing moderate-to-high agreement.