StagedWorkspace: A Versioned Workspace for Knowledg... | AI Research

StagedWorkspace: A Versioned Workspace for Knowledge-Work Agents addresses the problem of version drift in AI agents that perform knowledge work. When agents search, edit, and submit digital artifacts like documents, spreadsheets, and slides, they often rely on different versions of the same file. This paper proposes a "workspace-state contract" that ensures all views—parsed search results, native files, and review diffs—are explicitly tied to a specific version of the workspace.

The Workspace-State Contract

The authors argue that current agent interfaces often force a trade-off: artifact-only systems make searching difficult, while parsed-only systems lose critical layout and formula data. Furthermore, unversioned workspaces allow agents to modify files without creating a durable record for review. StagedWorkspace solves this by binding parsed records and review diffs to the content hashes of native files. As the agent modifies a file, the system updates the workspace state, invalidates stale parsed records, and triggers a refresh, ensuring the agent always operates on a synchronized version of the project.

How the System Works

The core of the approach is the SW-Agent, which implements three synchronized views:

• Native Workspace: The authoritative files used for execution and submission.

• Parsed Records: A searchable cache tagged by source path and content hash, allowing the agent to locate evidence without losing document structure.

• Journaled Review: A diff-based record that tracks changes between the starting workspace and the current state, allowing the agent to inspect edits before final submission.
  The system uses a hash-keyed synchronization mechanism. After a batch of tool calls, the harness scans the sandbox for changes. If a file’s hash has changed, the system marks the corresponding parsed records as stale and queues them for an update, preventing the agent from relying on outdated information.

Performance Results

In controlled ablations, the authors found that providing dual access to both parsed and native views significantly outperformed single-view configurations. On the OfficeQA Pro benchmark, dual access improved Pass@1 scores by 8.3–12.1 points. On the APEX-Agents benchmark, which involves complex, multi-format project folders, the dual-access approach improved mean rubric scores by 4.7–9.2 points.
When comparing SW-Agent against published results using the same models, the system showed notable gains. For example, using Gemini 3.1 Pro on OfficeQA, SW-Agent achieved 63.9% accuracy compared to a published score of 29.3%. On APEX-Agents, the GPT-5.4 Nano model reached 42.1% with SW-Agent, compared to a published score of 25.5%.

Considerations for Implementation

The authors identify workspace state as an experimental variable that influences agent performance. While the study demonstrates that synchronized views and visible diffs improve outcomes, the researchers note that their evaluation focused on specific benchmarks like OfficeQA Pro and APEX-Agents. The study also excludes certain tasks that depend on external APIs (such as specific EDGAR financial data) and notes that performance gains were measured against existing benchmark protocols. The findings suggest that future benchmarks for knowledge-work agents should treat evidence, staged edits, and submitted artifacts as explicit state transitions to better reflect the requirements of complex, multi-step tasks.