Source-linked AI summary
Working Through Representation: Malleable Workspaces for Human-Agent Knowledge Work
Nan Chen, Long Ling, Luna K. Qiu, Zilong Wang, Yuqing Yang
TL;DR
Conversational histories may not show how relevant materials currently relate or guide subsequent work. The paper develops the Malleable Workspace model and examines it through a 58-paper comparison and a three-month field study; systems varied in representational support, and the study refined the model’s account of connections to the Work Environment.
Problem
Conversation histories preserve exchanges but may not expose current relations among work materials or make them available for subsequent action.
Source: §1 · p. 2Method
The authors iteratively developed the Malleable Workspace through conceptual analysis, comparison of 58 papers, and a three-month MalleableSpace study with seven experienced agent users.
Source: §1 · p. 2Results
Across systems, system-assembled context was common, local incorporation outpaced reorganization, and fewer systems used represented organization to change distinct artifacts or environments.
Source: §1 · p. 2Takeaways & Limitations
The investigations refined the model’s account of how representation-grounded action depends on and can change the broader Work Environment.
Source: §1 · pp. 2–3The model’s generative and practical value as a design framework remains to be further examined.
Source: §6.3 · pp. 13–14
Abstract
from arXiv · showhide
Generative AI supports knowledge work, yet conversational histories may not capture the evolving organization that guides subsequent action. We introduce the Malleable Workspace, a conceptual model of how humans and agents form and revise a Work Representation and use its organization to advance work within a wider Work Environment. We developed it iteratively through conceptual analysis, comparison of 58 papers, and the construction and three-month longitudinal study of MalleableSpace with seven experienced agent users. Systems varied in what they represented as units and relations. System-assembled context was common; more systems supported local incorporation of agent outputs than reorganization, and fewer supported consequential agent-driven changes beyond the representation. The field study showed how representational roles and organization developed in use. These observations refined the model's account of how representation-grounded action depends on and can change the Work Environment. We contribute a conceptual and analytic foundation for representation-centered human--agent design.
1 Introduction
Conversational histories preserve interaction but do not necessarily expose how relevant materials relate or what those relations mean for next steps. The paper introduces and iteratively develops the Malleable Workspace model, then uses a cross-system analysis and field study to examine representation-grounded work.
1 Introduction
- Conversation histories retain exchanges, but may leave current relationships among materials unclear and require users to reconstruct context for subsequent action.This challenge motivates representing evolving work organization rather than relying on conversational progression alone.Source: §1 · p. 2
- The Malleable Workspace models how humans and agents form and revise a Work Representation that grounds action and connects to a broader Work Environment.The representation can draw on and selectively incorporate materials and outcomes from that environment.Source: §1 · p. 2
- The model was developed iteratively through conceptual analysis, comparison of 58 papers, and a three-month study of MalleableSpace with seven experienced agent users.Insights from the comparison and study prompted revisions to the model and renewed corpus analysis.Source: §1 · p. 2
- Across the corpus, systems varied in represented units and relations; system-assembled context was common, while agent-directed acquisition was less common.Local incorporation of agent outputs was more prevalent than reorganization, and fewer systems used represented organization to change distinct artifacts or environments.Source: §1 · p. 2
- The field study found that representational forms took on multiple roles and that their organization developed through use.Using organization beyond MalleableSpace initially required carrying context to existing tools and agents; connecting the Space to those agents enabled represented organization to guide changes.Source: §1 · p. 2
- The work contributes a conceptual model, an analysis of 58 papers, and an operationalization through MalleableSpace and its longitudinal study.The model’s account was refined to include how representation-grounded action depends on and can change the Work Environment.Source: §1 · p. 3
2 Related Work
Prior research addresses conversational continuity, external representations, and models of human–AI interaction. The paper builds on these strands with a model focused on how evolving represented work mediates collaboration and connects to subsequent action.
2.1 Conversation History and Continuity of Work
- Conversation histories support continuity and branching, yet work-relevant information can be distributed across turns and paths as users’ intentions evolve.Prior tools support reuse and inspection of retained information, while motivating ways to bring distributed information together and reorganize it for changing needs.Source: §2.1 · p. 3
2.2 External Representations in Knowledge Work
- External representations shape what information is available and which operations people can perform, while persistent organization can remain revisable over time.Prior accounts connect representational manipulation to understanding, computation, and maintaining constraints among organized objects.Source: §2.2 · p. 3
2.3 Interaction Models for Human–AI Work
- Human–AI interaction models describe instruction formation, reusable instruments, persistent generation objects, and collaborative coordination and adaptation.Other work examines analytical artifacts and shared workspaces that organize evidence and interpretations for further action.Source: §2.3 · pp. 3–4
- The Malleable Workspace models how evolving represented organization mediates human–agent collaboration and how action changes the broader Work Environment.It makes these relationships explicit to support analysis of the mediation and identify design opportunities.Source: §2.3 · p. 4
3 From Structural Mismatches to the Malleable Workspace
Conversation-centered interaction can mismatch the way knowledge work evolves: relevant elements, their relations, the current state, and subsequent actions may not align with turns and branches. The Malleable Workspace frames a revisable representation as a medium for organizing work and grounding action within a wider environment.
3.1 Structural Mismatches in Conversation-Centered AI Interaction
- Conversation-centered interaction creates unit, relation, state, and progression mismatches as work evolves across turns, sessions, and sources.Users may need to reconstruct relevant context to compare explanations, seek evidence, and revise a manuscript.Source: §3.1 · p. 4
3.2 The Malleable Workspace Model
- The model centers an evolving Work Representation that humans and agents revise to organize relevant elements and relations and ground subsequent action.Units are coherent work objects whose boundaries need not match source messages; relations express how units are currently understood.Source: §3.2 · p. 4
- The Work Representation selectively externalizes what matters from a Work Environment comprising tools, source materials, artifacts, agents, and other resources.Human guidance directs how the representation is used for an agent action without requiring context reconstruction in every prompt.Source: §3.2 · p. 4
- Representation-grounded action can be followed by interpreting outcomes and revising represented units and relations, or by changing the broader environment.An agent-generated comparison may qualify an explanation and prompt an update to the representation; using it to revise the actual manuscript is an environmental change.Source: §3.2 · pp. 4–5
4 The Malleable Workspace as an Analytic Lens
The paper applies the Malleable Workspace as a framework for comparing how systems represent work and connect it to agent action. Its analysis distinguishes representational forms, grounding, incorporation, and the reach of changes into the broader work environment.
4.1 Comparative Approach
- The comparison covers 61 implementation cases from 58 papers selected for inspectable, revisable representations used in subsequent agent action.Coders used an iterative shared codebook to record units, relations, grounding, incorporation, and changes to other work objects.Source: §4.1 · pp. 5–6
4.2 Forming Work Representations through Units and Relations
- Systems varied in addressable units and in the interpretive, compositional, and developmental continuity preserved through relations.Units made different aspects of work directly actionable, while relations retained meaning, grouping, or how work developed.Source: §4.2 · p. 6, §4.2.1 · p. 6, §4.2.2 · p. 6
- Representational relations preserved development by connecting derived ideas to sources or revisions to earlier versions.Other systems preserved interpretations through visible groupings or compositions through hierarchies and scoped relations.Source: §4.2.2 · p. 6
4.3 Grounding Agent Action in Represented Work
- System-assembled context was common, whereas agent-directed acquisition of additional material during action was less common.Systems could add context through structural scope rules or computational retrieval, ranking, filtering, clustering, and inference; agents in some systems also assessed intermediate results and sought further information.Source: §4.3 · pp. 6, 8
4.4 Incorporating Agent Outputs into the Work Representation
- Local incorporation of agent outputs was more widely supported than reorganization of existing units.Systems variously left outputs for human incorporation, presented candidates for approval, or incorporated them through predefined rules.Source: §4.4 · p. 8
4.5 Changing the Work Representation and the Broader Work Environment
- More systems supported agent-driven changes within represented work than changes to distinct artifacts or environments beyond it.Within-representation work included creating or revising artifacts and structures; exporting them for later use was distinct from using them to guide changes to another object.Source: §4.5 · p. 8
- Some systems used represented organization to guide changes to distinct artifacts or environments, including drafts, reports, slides, and live webpages.This differs from changing the representation itself because the represented organization guides action on another work object.Source: §4.5 · p. 8
5 MalleableSpace: From Instantiation to Model Refinement
The study examines how MalleableSpace users formed, revised, and used represented organization in self-selected work, and how that work connected to the wider Work Environment.
5 MalleableSpace: From Instantiation to Model Refinement
- The study investigated how participants formed, revised, and used represented organization to guide action in self-selected work.It also examined connections to resources and action in the wider Work Environment.Source: §5 · p. 8
- MalleableSpace was constructed as a general-purpose instantiation of the emerging Malleable Workspace model.Source: §5 · p. 8
- The researchers analyzed participant practices and design iterations to examine operational implications and clarify the model’s account of the Work Environment.Source: §5 · p. 8
5.1 MalleableSpace: Instantiating the Model
MalleableSpace instantiates the model as an editable visual workspace where users and agents organize units, provide context, and revise represented work.
5.1 MalleableSpace: Instantiating the Model
- An infinite visual canvas hosted heterogeneous units whose spatial arrangement, Frame membership, and labelled edges users could revise.Users could drag selected fragments from chats, PDFs, and notes onto the canvas for further organization and use.Source: §5.1 · p. 8
- Users could chat generally or open threads from persistent Agent Nodes positioned on the canvas.Source: §5.1 · p. 8
- Users supplied action-specific context, while Agent Nodes let agents inspect nearby, connected, or same-Frame units and search the Space for more materials.This combined user-selected inputs with structural assembly and agent-directed acquisition.Source: §5.1 · p. 9
- Agents could revise unit content and relations or reorganize existing units, and users could keep or revert those changes.The initial capabilities focused on work within the Space; connections to user-owned agents came later.Source: §5.1 · p. 9
5.2 Longitudinal Field Study
The three-month formative study followed seven experienced agent users as they incorporated MalleableSpace into self-selected work, with participant feedback informing system development.
5.2.1 Participants.
- Seven experienced agent users from software development, product management, and HCI/AI research took part and remained involved throughout the deployment.All used LLM-based agents daily; three had substantial experience with visual workspace tools.Source: §5.2.1 · p. 9
5.2.2 Study Procedure and System Iteration.
- Participants used the system in their own work without assigned tasks or minimum-use requirements, and joined weekly group reflections during the three-month deployment.The team added user-owned-agent support in stages, including ACP in Week 4 and progressively broader Space inspection and modification in Weeks 7–9.Source: §5.2.2 · p. 9
5.2.3 Data and Analysis.
- The qualitative corpus combined weekly reflections, exit interviews, group chat, and participant-selected Space cases; researchers iteratively coded materials and refined themes through team discussion and disconfirming cases.To protect privacy and ownership, researchers did not access or analyze other Space content.Source: §5.2.3 · p. 9
- Participants reported creating 43 Spaces, ranging from 3 to 11 per person; the largest contained 629 nodes.Reported uses included literature review, software development, presentation preparation, and personal knowledge work.Source: §5.2.3 · p. 9
5.3 Representational Roles and Evolving Organization
The field study found that participants assigned work-specific roles to represented units and relations, then developed organization through human judgment and agent-assisted revision.
5.3.1 Representational Roles Emerged in Use.
- Representational roles depended on how participants situated and used units and relations, not on node type alone.The findings extend the model by showing roles and represented organization developing through human judgment and agent-assisted revision.Source: §5.3.1 · p. 9, §5.3 · p. 9
- A Text node could preserve an unresolved decision and prompt agent discussion, while excerpts arranged around a transcript supported questions for subsequent conversations.Wording, visual treatment, placement, and relations distinguished materials to consult, questions to pursue, and guidance for action.Source: §5.3.1 · p. 10
- A Frame of layout guidelines served both as guidance for an agent and as organization for software issues, with colors marking status and edges showing dependencies.Source: §5.3.1 · p. 10
5.3.2 Organization Evolved Through Human-Guided Agent Assis-
- P5 treated an agent’s loose thematic grouping as a proposal, then worked with the agent to consolidate it into three broader themes while retaining comments for discussion.Source: §5.3.2 · p. 10
- Participants left materials loosely arranged during exploration, articulated structure when it emerged, and recombined selected elements from parallel slide versions with agent help.Agents materialized these decisions through changes across existing units and relations.Source: §5.3.2 · p. 11
5.4 Connecting Represented Work and the Work Environment
Connecting MalleableSpace to participants’ existing agents let represented work draw selectively on external resources and guide changes to external artifacts. This connection also motivated treating the Work Environment as a core part of the model.
5.4 Connecting Represented Work and the Work Environment
- Without connections to existing resources and agents, participants manually transferred context into and out of MalleableSpace to advance work.P1 re-uploaded papers and copied Zotero notes; P7 could not make a repository available to the in-Space agent and copied proposals to a coding agent.Source: §5.4 · p. 11
- Connecting MalleableSpace to user-owned agents let them draw on external resources while working with represented materials and judgments.This supported selective incorporation of external materials and requests for action on external artifacts without reproducing the whole work environment in the Space.Source: §5.4 · pp. 11–12
- Agents brought external materials and proposals into represented work for continued inspection and revision.Examples include organizing Zotero papers into thematic notes and comparing plans from Copilot and Codex using shared context.Source: §5.4 · p. 12
- Participants also used represented organization to guide changes to external resources and artifacts.P1 updated a Zotero library based on revision work, while P7 used organized GitHub issues and repository-grounded discussion to inform local code changes.Source: §5.4 · p. 12
- These observations prompted the model to include a Work Environment containing both resources that inform represented work and artifacts changed through representation-grounded action.The refinement concerned how represented work participates in activity beyond its own revision, not simply what the Space can contain.Source: §5.4 · p. 12
6 Discussion
The discussion distinguishes working on a representation by revising its organization from working through it when that organization guides action. It then derives design considerations and identifies limits to the model’s evaluation and empirical scope.
6.1.1 Working On and Through Representation.
- Working on a representation means forming and revising its organization; working through it means using that organization to guide subsequent action.Changing organization can itself develop understanding, while working through the representation can affect external artifacts; the two form a recursive loop, but visibility, generation, and representation do not guarantee grounding, incorporation, or consequence.Source: §6.1.1 · p. 12
6.1.2 Malleability in Continued Human–Agent Work.
- Malleability includes both local updates and coordinated revisions across units and relations as work develops.The analysis found local incorporation of agent outputs more common than reorganization; the study showed participants using agents to regroup, relate, and recombine existing units.Source: §6.1.2 · p. 12
- Under human guidance, agents can interpret materials, reorganize units and relations, and apply explicit organizing schemes to the representation.Guidance may come from reusable instructions or revisable notes in the Space; the model is about representation, action, and environment, not a particular interface category.Source: §6.1.2 · pp. 12–13
6.2.1 DC1. How Should Work Be Made Addressable through Units
- Designers should choose addressable work units and relations that balance stable schemas with inspectable, flexible organization.Agents can interpret heterogeneous sources into candidate units and relations, while designers consider extraction granularity and how units are represented and modified.Source: §6.2.1 · p. 13
6.2.2 DC2. How Should a Work Representation Ground Particular
- Designers should determine how human guidance, rules, and agent information seeking shape the context grounding each action.Broader context may retain relevant connections but add irrelevant material; narrower context improves predictability but may omit information needed to interpret or act.Source: §6.2.2 · p. 13
6.2 Design Considerations for Malleable Workspaces
- Designers should decide whether agent outputs remain transient or are incorporated as local updates or broader reorganizations, and who approves changes.Local updates offer predictability; broader reorganization may better reflect developing understanding but risks disrupting established structure.Source: §6.2.3 · p. 13
- Designers should connect represented work to external resources and artifacts in both directions, while controlling and inspecting what crosses the boundary.Without such connections, participants became context carriers manually transferring information into and out of the Space.Source: §6.2.4 · p. 13
6.3 Limitations and Future Work
- The model’s practical design value remains to be tested, and the three-month study’s seven experienced users covered a relatively narrow range of users and work contexts.Future work is proposed to examine independent application of the model and broaden empirical evidence across populations, occupations, tasks, and settings.Source: §6.3 · pp. 13–14
7 Conclusion
The paper presents the evolving representation as an alternative to interaction organized only through exchanges, with its organization guiding what happens next. Its analysis and longitudinal study support a model connecting representation, human and agent action, and the broader Work Environment.
7 Conclusion
- The Malleable Workspace frames an evolving representation as a revisable medium that organizes work and grounds human and agent action in a broader Work Environment.The cross-system analysis demonstrated comparative and diagnostic value, while the MalleableSpace study elaborated and refined the model through sustained use.Source: §7 · p. 14