Staged Workspace architecture links file states to content chunking and doubles the accuracy of knowledge agents

AI models and software agents are rapidly moving into knowledge work, producing and modifying persistent digital artifacts such as code repositories, text documents, spreadsheets, presentations, reports, and interactive computational notebooks. These systems face a structural gap caused by the lack of precise data synchronization: models search parsed text representations, edit raw files, review change diffs, and deliver final artifacts, with each step potentially operating on an entirely different version of the workspace without a mechanism to enforce stage consistency.

The paper introduces the concept of workspace state contracts and proposes the Staged Workspace architecture to manage file evolution.Developed by researchers Yining Hua, Hongbin Na, Yifan Zhu, Akshay Kalose, Cyrus Ayoubsha, and Levi Lian, the framework addresses limitations found outside conventional software engineering. While coding agents rely on clear in-repository contracts to manage retrieval, diffing, and test execution, work across PDF files, spreadsheets, slides, and mixed-format folders lacks comparable contracts that explicitly tie each visual view to the evolving workspace state.

The technical architecture of Staged Workspace links parsed records and review diffs to cryptographic hashes of the underlying raw files with every modification. This mechanism provides dual-modality access, allowing an agent to work with both parsed representations and raw files simultaneously, which yielded the highest scores across all evaluated models on the OfficeQA Pro and Apex Agents platforms. Dual-modality access improved pass@1 success rates on OfficeQA by 8.3 to 12.1 percentage points over restricted single-view baselines, while lifting average benchmark scores on Apex by 4.7 to 9.2 points.

Empirical benchmarks demonstrate that the architecture can double the effectiveness of language models when paired with a disciplined staged workspace.SW-Agent achieved a 63.9 percent success rate when powered by Gemini 3.1 Pro on OfficeQA, substantially outperforming the published baseline of 29.3 percent for the same model. The agent also scored 42.1 points with GPT-5.4 Nano on Apex, compared to its published score of 25.5 points. Parallel experiments focused on review mechanisms across 57 file-editing tasks confirmed that giving agents visibility into change diffs markedly improves observed performance scores.

These findings demonstrate that workspace state is a primary empirical variable in knowledge-work agent development, opening the way for new evaluation benchmarks that treat evidence gathering, staged edits, and committed artifacts as explicit, recorded state transitions.