Source-linked AI summary
Ream: Unfolding Mutual Awareness in Human-Agent Workspaces
Peiling Jiang, Sangho Suh, Varsha Kishore, Jonathan Bragg, Haijun Xia, Pao Siangliulue, Daniel S. Weld, Amy X. Zhang, Joseph Chee Chang
TL;DR
Agents can act faster than people can monitor in shared workspaces, while users’ evolving interests may not appear in chat. Ream makes both parties’ engagement with literature-review documents visible and retrievable; researchers used these traces to inspect and steer agent work, and shared histories supported coordination.
Problem
In shared workspaces, agents’ actions can outpace human monitoring, and files and chat do not capture users’ evolving engagement and interests over time.
Source: §1 · p. 2Method
Ream tracks and visualizes users’ and agents’ engagement in shared literature-review artifacts, making the history available to both parties.
Source: §1 · p. 2Results
In studies with eighteen researchers, embedded awareness signals helped participants understand, inspect, and steer agent work, recognize latent interests, and supported agents’ implicit coordination through shared histories.
Source: §9 · p. 11Takeaways & Limitations
The findings offer a reusable design approach for extending mutual-awareness principles from human-human collaboration to human-agent workspaces.
Source: §9 · p. 11The studies involved graduate researchers who frequently used coding agents; broader recruitment could extend the findings to novice users and other research roles.
Source: §7.2.6 · p. 10
Abstract
from arXiv · showhide
As AI agents work alongside humans in shared workspaces, a mutual awareness challenge arises: agents act at speeds that outpace human monitoring, and users' evolving interests are not always expressed in chat. This challenge is especially pressing in literature review, where both parties retrieve, read, and synthesize a growing body of papers. We present Ream, a literature review workspace that supports mutual awareness through structured artifacts, bidirectional engagement tracking, and localized visualizations. Users can see each party's activity within these documents, and agents can retrieve the same history to guide their work. In studies with eighteen researchers, participants used these traces to inspect evidence, steer agents, communicate through annotations, and reflect on their research focus. Shared histories also helped agents build on earlier work. These findings inform how engagement traces within shared documents can support transparency, personalized assistance, and coordination in human-agent knowledge work.
1 Introduction
In shared workspaces, agents’ activity can outpace human monitoring while users’ evolving interests remain implicit in their workspace activity. Ream tracks and visualizes both parties’ engagement in shared literature-review artifacts; studies show how researchers used these traces to inspect and steer work and reflect on their focus.
1 Introduction
- Existing chat interfaces make it difficult to map agents’ long action traces to affected files, while agents often lack a history of users’ engagement and changing priorities.Files capture their current state, not how they came to be or how users have engaged with them.Source: §1 · p. 2
- Literature review makes mutual awareness especially relevant because researchers assess claims against sources while questions and interpretations evolve across iterative cycles of searching, reading, organizing, and synthesis.Researchers also revisit, annotate, and edit papers even when agents can retrieve and summarize them.Source: §1 · p. 2
- Ream represents literature review with shared paper, collection, and synthesis artifacts, tracking and visualizing both users’ and agents’ actions across documents.Localized engagement signals help users follow agents’ efforts and let agents use users’ workspace activity to align assistance.Source: §1 · p. 2
- Researchers used workspace traces to inspect evidence, steer assistance, reflect on their research focus, and carry context across multiple agents.These patterns recurred in the longitudinal study and informed design implications for ongoing collaboration.Source: §1 · p. 2
2 Background and Related Work
Prior collaboration research highlights persistent traces as a basis for awareness, coordination, and shared understanding. Ream applies these ideas to human-agent workspaces, where agents need to track user activity and users need ways to inspect and steer evolving literature-review work.
2.1 Awareness in Human-Human Collaborative Workspaces
- Human-human workspace-awareness research frames awareness around who is working on what, where collaborators have been, and what actions they have taken.Prior systems used persistent traces on shared artifacts to support coordination, including without explicit communication.Source: §2.1 · p. 3
2.2 Awareness in Human-Agent Collaborative Workspaces
- Human-agent collaboration requires visibility into what agents do and why, while Ream also lets agents retrieve users’ activity histories to follow their work beyond conversation.Users can see papers retrieved, questions asked, and passages extracted; agents can see papers revisited and notes added.Source: §2.2 · p. 3
- Ream uses accumulating workspace attention—not just conversation—to infer evolving interests and suggest tasks aligned with what users care about.Signals include closely read papers, notes, and revisited summaries.Source: §2.2 · p. 3
2.3 AI-Supported Literature Review Tools
- Current deep-research tools often generate a single report per query and lack a shared, evolving space for users to inspect and steer the review process.Ream instead embeds agents in a workspace of evolving papers, curated collections, and synthesis reports that both parties can inspect and act on.Source: §2.3 · pp. 3–4
3 Design Goals
Ream’s design goals address work distributed across documents and attention that shifts over time. They call for shared domain-specific artifacts, engagement histories for both parties, and localized visualizations that agents can retrieve.
3 Design Goals
- Represent the literature-review workflow with shared domain-specific artifacts, including papers, collections, and synthesis reports, so both parties can inspect and build on intermediate work.These artifacts provide concrete surfaces on which awareness can be visualized.Source: §3 · p. 4
- Track opens, edits, stars, and notes over time because a workspace snapshot cannot show agents’ accumulated effort or users’ evolving interests.Engagement histories let both parties draw on activity beyond which files currently exist or are open.Source: §3 · p. 4
- Embed attention visualizations in file trees and documents, and make the same engagement signals retrievable by agents to guide their retrieval, organization, and synthesis.Interaction histories record who acted, what they did, which document they worked on, and when.Source: §3 · p. 4
4 Agentic Literature Review with Ream
Ream provides a shared workspace where researchers and agents work with structured literature-review documents. Its paper files, collections, and reports serve as artifacts to inspect, organize, and develop the review together.
4 Agentic Literature Review with Ream
- Ream is a web-based literature-review workspace with a file tree, a central document viewer and editor, and a chat panel for managing and interacting with agents.Multiple workspaces and parallel agents support separate or concurrent searching and synthesis.Source: §4 · p. 4
4.1 Representing the Literature Review Workflow with Structured Artifacts (DG1)
- Structured paper, collection, and report files give researchers and agents concrete surfaces to inspect, annotate, and build on, while grounding awareness tracking and visualization.The artifacts represent stages of the literature-review workflow.Source: §4 · p. 4, §4.1 · p. 4
- Each .paper file consolidates paper information, full text, PDFs, researcher notes, extracted quotes, and review-specific annotations into one artifact.Users can add notes; agents can extract evidence quotes and retrieve metadata and prior explorations, making each party’s engagement visible to the other.Source: §4.1 · p. 5
- Automatically created .papers files preserve the results of each search, while curated collections let users organize papers by theme or task and agents reference a collection as a shortcut.The same underlying paper file is shared across workspace locations, so notes and annotations appear everywhere it is represented.Source: §4.1 · p. 5
- Editable .report files hold literature syntheses with clickable links to underlying paper files, and can be drafted as outlines for users to reshape before detailed writing.Users can maintain multiple reports to explore different framings without overwriting earlier syntheses.Source: §4.1 · p. 5
5 Unfolding Bidirectional Awareness
Ream records user and agent engagement with shared artifacts over time, surfaces activity within the workspace, and lets agents query the same history to inform their work.
5.1 Tracking Engagement Over Time (DG2)
- Each signal records when and by whom an action occurred, its category, affected files, and strength; heuristics reduce incidental activity rather than capture every interaction exactly.Repeated opens and hovers are limited to once per 30 seconds, while sustained focus and stronger actions receive higher signal values.Source: §5.1 · pp. 5–6
5.2 Surfacing Awareness to Users with Localized Visualizations (DG3)
- Ream embeds awareness visualizations in workspace artifacts instead of placing them in a separate dashboard.The file tree uses separate user and agent bars, scaled independently across visible files, to show engagement across the collection.Source: §5.2 · p. 6, File Tree Attention Bars · p. 6
File Tree Attention Bars.
- Per-file histories show which actor performed which actions and when, providing a chronological view of engagement with each artifact.Users can check, for example, when an agent last edited a report and whether it recorded read or write signals.Source: Per-File Signal History · p. 6
PDF Read Wear.
- The PDF reader highlights source passages matched to quotes returned by the agent’s full-text exploration tool, so users can inspect evidence in context.The system fuzzymatches quotes to sentence-level bounding boxes extracted by a Grobid-based parser.Source: PDF Read Wear · p. 6
Inline Reference Chips.
- Clickable reference chips show user and agent engagement with cited papers inside reports and paper notes, and open those papers for closer inspection.This lets users check engagement without returning to the file tree.Source: PDF Read Wear · p. 6
5.3 Surfacing Awareness to the Agent (DG3)
- Agents can query engagement history to retrieve recently read or frequently revisited papers, signals within an artifact, or filtered chronological events.The four read_signals modes support different targets and time windows.Source: §5.3 · p. 6
- When users focus the chat input, Ream generates task suggestions from workspace activity, such as incorporating recent notes or flagging a highly cited unread paper.Figure 3 depicts suggestions appearing when the chat input is focused.Source: §5.3 · pp. 6–7, Figure 3 · p. 6
6 Example Workflow with Ream
The example follows Anna as she uses Ream to gather, curate, and synthesize literature, while engagement traces and suggestions help her inspect and steer the work.
6 Example Workflow with Ream
- Anna asks the agent to search across three themes from her introduction, then receives four organized collections populated with papers from search results.She inspects agent attention and relevance questions, and can check why papers were archived.Source: §6 · p. 7
- Anna trims broad survey papers, stars a relevant feature-explanations paper, and adds a note challenging a claim about feature weights.Opening papers, reading PDFs, writing notes, starring, and removing references update localized attention signals.Source: §6 · p. 7
- The agent uses Anna’s engagement to weight its synthesis: it gives more detail to papers she engaged with and excludes papers she removed.Inline reference bars help her navigate to papers and verify the work.Source: §6 · p. 7
- Suggestions based on Anna’s recent activity shift as she reads and annotates, prompting report updates and next steps tied to specific papers.One suggestion incorporates recent notes; a later one asks her to list regulatory requirements from a paper after she focuses on that section.Source: §6 · p. 7
7 Evaluation
Across lab and week-long use, researchers used Ream’s shared activity traces to inspect and steer agent work, reflect on their focus, and organize review workflows. The longitudinal study found recurring patterns of implicit steering and continuity across sessions.
7.2 Findings
- 10 of 12 lab participants asked unprompted to continue using Ream or export their reports; sessions averaged 181 papers and 11 messages.The lab study examined a continuous work session, while the longitudinal study involved six researchers using Ream for one week.Source: §7.2 · p. 8, §7 · p. 8, §7.3 · p. 10
7.2.1 Awareness visualizations help users inspect and steer agent work (RQ1).
- Ten participants reported clearer understanding of agent activity, using traces to inspect synthesis evidence and decide what to read or request next.Participants followed source passages to verify claims, and found workspace visualizations easier to comprehend than lengthy chat traces.Source: §7.2.1 · p. 8
- Participants used engagement cues to follow agent emphasis or examine overlooked papers, and one redirected retrieval when a priority subfolder received little attention.P8 intervened while the agent was working and issued a targeted follow-up prompt.Source: §7.2.1 · p. 8
7.2.2 Agents respond to researchers’ interests through workspace activity (RQ2).
- Participants noticed agents responding to workspace activity beyond chat, including report annotations that informed later searches and surfaced relevant topics.P6 said integrated workspace context avoided having to reformulate a query and re-provide context.Source: §7.2.1 · p. 9
7.2.3 Visualizing their own activity helps researchers recognize their evolving focus (RQ1, RQ2).
- Seeing their own traces helped participants notice unintentional patterns in their reading and recognize how their research focus was developing.P6 noticed repeated returns to a paper across reports and eventually starred it.Source: §7.2.3 · p. 9
7.2.4 Researchers adapt awareness visualizations to their review strategies (RQ3).
- Researchers adapted visualizations to their review strategies, using engagement bars to identify core papers, explore less-engaged sources, and check report coverage.P12 used the file tree to confirm that a report relied heavily on one source; P1 checked papers the agent had not chosen.Source: §7.2.4 · p. 9
7.2.5 Shared artifacts support flexible literature review workflows (RQ3).
- Reports served as navigational and working surfaces, while participants used inline links to verify claims and activity traces revealed varied workflows.P7 gathered and synthesized before reading; P9 interleaved searching, synthesis, organization, and reading.Source: §7.2.5 · p. 9, Figure 5 · p. 9
7.2.6 Study Scope and Limitations.
- The lab study had no baseline or ablation, and participants were graduate researchers who frequently used coding agents, limiting the sample’s breadth.The authors identify novice users and other research roles as targets for broader recruitment.Source: §7.2.6 · p. 10
8 Discussion and Future Work
Ream’s localized traces helped researchers inspect agent work at passages, papers, and collection levels. The discussion considers extending this awareness approach to research support, other workflows, multi-agent collaboration, and proactive assistance.
8 Discussion and Future Work
- Passage highlights and collection comparisons let participants verify claims and redirect the agent while reading and writing.These uses illustrate awareness at passage, paper, and collection levels.Source: §8 · p. 10
8.1 Unfolding Awareness for Research Support beyond Literature Review
- Engagement signals can encode evolving interests beyond literature review, including identified gaps and research threads researchers have deprioritized.The authors suggest this history could support framing contributions and drafting related work.Source: §8.1 · p. 10
- Cumulative activity bars, timestamped histories, and notes help collaborators interpret patterns without inferring the purpose of every interaction.A single open may reflect distraction, while repeated visits may reflect confusion.Source: §8.1 · p. 10
8.2 Unfolding Awareness for Other Domain-Specific Agentic Workflows
- Shared, structured artifacts and localized traces could transfer to legal due diligence and qualitative research workflows.Examples include contracts, clause collections, and assessment briefs, or transcripts, codebooks, and thematic memos.Source: §8.2 · p. 10
- Connecting traces to questions and source passages lets users inspect how a result was reached and decide whether to rely on a claim.The authors note that evaluations varying output accuracy could examine how evidence inspection shapes confidence and reliance.Source: §8.2 · p. 10
- Selected-action traces link signals to artifacts for retrieval and visualization, while complete session replays could retain additional context.The authors note that models could identify relevant actions and artifacts from recordings as capture capabilities develop.Source: §8.2 · p. 11
8.3 Unfolding Awareness in Multi-Agent Collaboration
- Researchers used multiple agents with varied strategies, and one second agent queried prior signals to begin complementary work.Participants also wanted to distinguish individual agents’ contributions when assigning follow-up tasks or tracing edits.Source: §8.3 · p. 11
8.4 Unfolding Awareness for Proactive Agent Assistance
- Suggestions beside the chat input let researchers choose when to take up agent initiative, while more assertive interventions raise the question of when to interrupt.The authors propose examining how suggestion timing and prominence affect use across research goals and activities.Source: §8.4 · p. 11
9 Conclusion
Ream uses structured artifacts, bidirectional engagement tracking, and localized visualizations to unfold mutual awareness between users and agents. Studies with eighteen researchers found that these traces supported inspecting and steering agent work, reflecting on interests, and implicit coordination.
9 Conclusion
- Studies with eighteen researchers found that embedded awareness signals helped participants understand, inspect, and steer agent work.The signals appeared within documents and file structures where users already work.Source: §9 · p. 11
- Visualizations also helped users recognize their latent interests, while shared signal history enabled agents to coordinate implicitly.Source: §9 · p. 11
- The findings extend mutual-awareness principles from human-human collaboration to human-agent workspaces and offer a reusable design approach across domains.Source: §9 · p. 11
A Implementation Details
Ream combines a collaborative web application stack with agent tools for literature review. It also links agent evidence quotes to sentence-level locations in PDFs for read-wear visualization.
A Implementation Details
- Ream uses React and Elysia with Bun, TypeScript, PostgreSQL, and Yjs for real-time collaborative editing.Agent and user edits share the Yjs layer, with character-level attribution in revision history.Source: A Implementation Details · p. 14
- The Ream agent uses the Pi agent SDK and Semantic Scholar APIs for paper metadata and full-text retrieval.Users select gpt-6-astra or gpt-6.1-sol for main agents; auxiliary processes default to gpt-6.1-sol.Source: A Implementation Details · p. 14
- For PDF read-wear, verbatim evidence quotes are fuzzy-matched to sentence-level bounding boxes extracted by a Grobid-based parser.Source: A Implementation Details · p. 14
B Action Signal Strengths
Action weights summarize recorded interaction as signed engagement scores, and activity bars display relative activity for each actor across displayed files.
B Action Signal Strengths
- Table 2 lists heuristic action weights that summarize interaction rather than measure attention or intent.Positive values add to the signed engagement score; negative values encode actions such as removing or downvoting.Source: B Action Signal Strengths · p. 14, Table 2 · p. 14
- For each artifact and actor, the activity bar sums positive weights and applies square-root normalization against that actor’s maximum among displayed files.Nonzero values have a minimum visible proportion of 0.1.Source: B Action Signal Strengths · p. 14
C Activity Traces for User Study Sessions
Figure 6 presents activity traces for users and agents, organized by action type during user study sessions.
C Activity Traces for User Study Sessions
- Figure 6 presents activity traces for users and agents, and for different action types in user study sessions.Source: Figure 6 · p. 15