StagedWorkspace: A Versioned Workspace for Knowledge-Work Agents
| Source: arXiv AI
Tags: AI agents, knowledge work, document editing, OfficeQA, APEX, agent infrastructure
StagedWorkspace gives AI agents explicit version tracking across parsed views, native files, and diffs—boosting OfficeQA Pass@1 by 8–12 points and APEX rubric scores by 4–9 points, doubling same-model performance on knowledge-work benchmarks.
Details
AI agents increasingly edit code repositories, documents, spreadsheets, and slides—but they often read one version of a file, edit a cached version, and submit yet another. This workspace-state mismatch silently degrades agent performance. StagedWorkspace formalizes a workspace-state contract: every view an agent sees must be explicitly tied to a specific content hash of the native file at that moment. This is analogous to the repository contracts in coding agents (diffs, search, tests), but applied to knowledge-work artifacts like PDFs, spreadsheets, and notebooks. The practical gains are substantial. Dual access to both parsed records and native files improves OfficeQA Pass@1 by 8.3–12.1 points and APEX mean rubric score by 4.7–9.2 points. SW-AGENT with Gemini 3.1 Pro scores 63.9% on OfficeQA vs a 29.3% published same-model baseline. With GPT-5.4 Nano on APEX, it achieves 42.1 vs 25.5 published—approximately doubling performance using the same model. The paper frames workspace state as an overlooked experimental variable in knowledge-work agent evaluation, suggesting published benchmarks conflate view inconsistencies with model limitations.