代理基礎設施
StagedWorkspace Uses Content Hashes to Synchronize Agents’ Parsed Views and Native Files, Boosting OfficeQA Pass@1 by 8.3 to 12.1 Points
The new system binds a document agent’s search results, native files, diff views, and final submissions to explicit versions, preventing the agent from reading stale parsed content while editing a newer file. Ablations within a fixed framework show that providing both parsed and native views outperforms either view alone, though the full codebase and cross-framework reproductions are still needed.
Software agents can typically rely on Git commits, diffs, and test results to verify the current state. But when handling PDFs, spreadsheets, presentations, or mixed-document workflows, the search index, preview, native file, and version submitted for evaluation may fall out of sync. [StagedWorkspace](https://arxiv.org/abs/2608.18050), published on August 18, defines this problem as a “workspace state contract”: every parsed record and diff visible to the agent must point to a specific hash of the native content.
Its workspace maintains baseline, staged, and committed versions. Whenever a native file changes, the system reparses it and updates the searchable records. The agent can use structured text for retrieval while also inspecting the original PDF, workbook, or slide deck, reducing the risk of missing layout, formulas, and chart information lost by the parser. A pre-submission diff then lets the agent inspect what it actually changed instead of relying solely on an earlier textual summary.
Within a fixed agent framework, the researchers separately removed either the native or parsed view. The dual-view configuration achieved the highest point estimate across all tested models. Compared with the more constrained single-view settings, it increased OfficeQA Pro Pass@1 by 8.3 to 12.1 percentage points and raised the average APEX-Agents rubric score by 4.7 to 9.2 points. The complete SW-Agent system achieved 63.9% on OfficeQA with Gemini 3.1 Pro and 42.1 on APEX with GPT-5.4 Nano; the authors cite previously reported results of 29.3% and 25.5, respectively, for the same models. In a separate set of 57 file-editing tasks, the configuration that displayed diffs also received higher observation scores.
The results remind developers that a document agent’s “context” consists of more than tokens: it also includes external state with version semantics. The [official OfficeQA toolkit](https://github.com/databricks/officeqa) provides original PDFs, parsed documents, and evaluation code, illustrating the engineering burden of maintaining multiple representation layers. However, cross-system scores depend on the model, tools, and execution framework, so the gap between the complete system and published baselines cannot be attributed entirely to version control. The most credible evidence currently remains the ablations conducted within the same framework. The next step is to test whether the content-hash contract continues to hold under collaborative agents, concurrent editing, parsing failures, and rollbacks.