Not yet human-reviewed
This source record was written automatically by the research pipeline and has not yet been reviewed by the maintainer or integrated into a topic page. Treat it as a provisional summary and verify claims against the original source.
StagedWorkspace: A Versioned Workspace for Knowledge-Work Agents
Citation. Hua, Y., Na, H., Zhou, Y., Kalose, A., Ayubcha, C., & Lian, L. (2026). StagedWorkspace: A Versioned Workspace for Knowledge-Work Agents. arXiv:2608.18050.
Summary. This preprint presents StagedWorkspace, an agent workspace that binds parsed search records, native artifact operations, review diffs, and submitted deliverables to explicit file-version state. In fixed-harness studies on OFFICEQA PRO and APEX-AGENTS, the authors report that synchronized parsed and native views improve benchmark performance relative to the more limiting single-view condition, while visible diffs improve scores on a small file-editing slice. The results concern the evaluated agent harnesses and task settings, not a general causal estimate for all knowledge-work systems.
Key claims
- The proposed workspace-state contract uses content hashes to tie parsed records and review diffs to the native files that an agent edits and submits, marking parsed records stale when a source file changes until re-parsing completes.
- Across three tested models on OFFICEQA PRO, dual parsed/native access improved Pass@1 by 8.3–12.1 points over artifact-only access in paired ablations; on APEX-AGENTS, it improved mean rubric score by 4.7–9.2 points over parsed-only access.
- In a paired review-axis ablation on 57 APEX file-editing tasks, the paper reports higher observed scores when agents could inspect tracked diffs before submission.
Evidence & limitations
- The paper evaluates a ReAct-style SW-AGENT under a fixed 250-tool-call budget on OFFICEQA PRO and APEX-AGENTS, using paired ablations that hold the model, parser, retriever, grader, tracker, and tool budget fixed within the stated comparisons.
- Published benchmark comparisons are contextual rather than one-to-one because harnesses, corpora, attempt budgets, retrieval backends, and parser choices differ; trajectory coding is illustrative rather than exhaustive, and benchmark graders score final outputs rather than intermediate state transitions.
- The Drive PDF exactly matches the public arXiv PDF by SHA-256; this is a preprint, not peer-reviewed evidence.