Revising an AI's Categories
When an AI organizes project notes, reconsidering the categories is a different task from improving the information filed under them.
In a reported workspace study, a participant asked an AI agent to group interview findings. Finding its categories too loose, they worked together to revise them into 3 broader themes, retaining comments for discussion.
The authors distinguish working on a project’s arrangement from working through it when that arrangement guides the next action. For everyday AI work, the difference matters before asking an agent to continue. Better summaries under the same headings leave those divisions intact. If the divisions no longer serve your question, the organization itself needs attention.
The October 22 preprint describes MalleableSpace, an editable workspace studied with 7 experienced agent users over 3 months. Researchers analyzed discussions and selected cases, not comparative work quality.
Leaving the scheme unfinished
Another participant abandoned a branching arrangement during exploration, then gave an agent a clearer scheme when one emerged. If the distinctions you care about are still changing, specifying a comprehensive scheme could demand a decision you are not ready to make.
Consider a hypothetical software team organizing bug reports. Grouping by component could help assign owners. Grouping by the customer task disrupted could help frame a different investigation: reports from payments and notifications might belong together when examining a failed purchase. That connection would be a reason to investigate, not proof of a shared cause or a reason to schedule every report in the same release.
The choice changes what further work would make sense. Asking for more detail within each component would not explore the possible connection across components. Asking for a shared explanation would not settle who should fix each part. The same reports could support either task, provided the team keeps the purpose of the arrangement clear.
What another view does not settle
An older VisCanvas comparison gave 20 graduate computer science students 20-minute chart-exploration tasks, with matching editing and suggestion tools in its spatial and chat interfaces. Counts of charts and data-column combinations did not significantly differ. That does not establish equivalence, and neither count measures the quality of the analysis.
In a separate 5-day CanvasConvo deployment with 24 people experienced with AI, some participants stayed with chat habits while others used a spatial view selectively. One was unsure what context a conversation branch inherited. Visible structure did not automatically make the use of information clear.
For the hypothetical team, either grouping could live in its existing tracker or an ordinary outline. Adding a workspace would mean deciding what an agent should take from it and reconciling it with materials already in use. File access might make an edit possible without showing that the grouping helped.
A useful arrangement would have to earn its maintenance. A report moved into a new group might still need its existing owner, and a suspected connection might deserve discussion before it becomes a work plan. Those are different judgments. An agent can be asked to change the categories without being asked to settle everything that follows from them.
Sources
- Nan Chen, Long Ling, Luna K. Qiu, Zilong Wang and Yuqing Yang, Working Through Representation: Malleable Workspaces for Human–Agent Knowledge Work, arXiv preprint, October 22, 2026, version 1.
- Yuki Ueno, Bretho Danzy III, Zhuojun Jiang and Chris Bryan, VisCanvas: A Node-Based Interface for Exploratory Visualization Authoring with LLMs, version 3, September 22, 2026.
- Rifat Mehreen Amin, Alperen Adatepe, Daniela Fernandes, Daniel Buschek and Andreas Butz, Conversations in Space: Non-Linear LLM Interaction in Everyday Use, arXiv preprint, version 2, September 15, 2026.