StagedWorkspace: Version Control for Knowledge-Work AI Agents
A Feynman-style deep reading of *StagedWorkspace: A Versioned Workspace for Knowledge-Work Agents* (Hua et al., Harvard University & Raycaster AI, arXiv preprint arXiv:2608.18050).
The Problem
Knowledge-work agents produce and modify persistent digital artifacts — codebases, documents, spreadsheets, slides. Unlike coding agents (which benefit from Git, plain-text source, and test-driven verification), non-coding knowledge work suffers from:
- Parsing difficulty: complex PDF layouts, Excel formulas, visual PowerPoint content
- Weak version control: most office formats lack mature state management
- Human-only validation: no deterministic tests for correctness
- W_t: authoritative native workspace (PDF, Excel, PPT) used for execution and submission
- C_t: parsed records tagged by source path and content hash — current if hashes match W_t, stale otherwise
- Δ_t = δ(W_0, W_t): format-specific diffs (text lines, spreadsheet cells, slide-level) against the initial workspace
- OfficeQA Pro: numeric QA on US Treasury Bulletins; 104 document-only + 29 web-evidence questions over a shared workspace of ~697 PDFs (1939–2025); metric is exact-match Pass@1
- APEX-Agents: 480 scored tasks across 33 professional worlds (management consulting, investment banking, law); ~166 mixed-format files per task; metrics are Pass@1 and mean rubric score
- Read axis: dual parsed/native access gives the highest point estimates for every tested model; artifact-only loses 8.3–12.1 points on OfficeQA and 4.7–9.2 on APEX rubric scores; parsed-only is middling
- Review axis: on 57 paired file-edit tasks, scores are higher when diffs (
workspace_diff,workspace_file_diff) are visible before submission - Hua, Y., Na, H., Zhou, Y., Kalose, A., Ayubcha, C., & Lian, L. (2026). StagedWorkspace: A Versioned Workspace for Knowledge-Work Agents. *arXiv preprint arXiv:2608.18050*
- Hua, Y., et al. (2025). OfficeQA: Benchmarking Knowledge Work in Realistic Office Workflows. *ICML 2025*
- Hua, Y., et al. (2025). APEX-Agents: A Benchmark for Autonomous Professional Agents. *NeurIPS 2025*
- Chen, M., et al. (2021). Evaluating Large Language Models Trained on Code. *arXiv:2107.03374*
The paper's core insight, formalized as the workspace-state contract: *every view must be explicitly bound to a specific version of the evolving workspace state.* Agents can violate this in three ways:
1. Artifact-only systems — keep native files, but search is hard (page-by-page browsing, irrelevant context) 2. Parsed-only systems — searchable, but lose layout, formulas, visual evidence; edits target an abstraction, not the native file 3. Unversioned mutable workspaces — agents overwrite/move/delete files with no journaled diff, making "what did I change?" unanswerable
The StagedWorkspace Architecture
Three synchronized views:
After each mutation tool batch, the system hash-scans changed files, advances W_t, marks affected parsed records stale, and asynchronously refreshes them. Key guarantee — the write-after-read contract: after edits, native operations resolve against the updated workspace, and parsed results either match that version or are flagged stale until refresh.
Experimental Results
Benchmarks:
Full system (same-model comparisons):
| Model | Benchmark | SW-Agent | Published | Gain | |---|---|---|---|---| | Gemini 3.1 Pro | OfficeQA | 63.9% | 29.3% | +34.6 pts | | GPT-5.4 Nano | APEX | 42.1 | 25.5 | +16.6 |
Ablations:
Why It Matters
1. Workspace state is an experimental variable — as important as model choice or prompting; agent research should report workspace management strategies 2. Dual views are complementary, not competing — the optimum maintains both, synchronized 3. Reviewability drives performance — visible diffs act like a human checklist effect 4. Benchmarks should score evidence, staged edits, submitted artifacts, and explicit state transitions
The bigger picture: AI systems are evolving from stateless functions to stateful participants. Persistent workspace modification raises demands for traceability, reviewability, rollback, and human authorization.