StagedWorkspace: A Versioned Workspace for Knowledge-Work Agents
> *"Chaos isn't a pit. Chaos is a ladder."* — adapted from *Game of Thrones*
---
The Core Problem
A knowledge worker — say, a management consultant — opens a directory containing 37 PDFs, 12 Excel sheets, 8 PowerPoints, and 5 Word documents. They ask an AI assistant to search for Q3 2025 market-share data, update revenue forecasts, review the changes, and submit a final integrated PPT. Hidden inside this flow is a fatal inconsistency: the file version the agent searches, the version it edits, and the version it submits can all be different.
The paper's central claim is that workspace state itself is an experimental variable for knowledge-work agents, and must be explicitly controlled.
---
Knowledge Work vs. Text Tasks
LLM-agent research has shifted across three generations:
1. Isolated text tasks — QA, summarization, translation; stateless. 2. Tool use — search, compute, code execution; still "one-shot." 3. Knowledge work — producing and modifying *persistent* digital artifacts (codebases, documents, spreadsheets, slides), evaluated by the work product itself.
Coding agents already benefit from plain text, mature Git-based versioning, test-driven verification, and deterministic execution. Non-coding knowledge work has none of this:
| File type | Parse difficulty | Version control | Verification | |---|---|---|---| | PDF | High (complex layouts) | Usually none | Manual | | PowerPoint | High (visual) | Limited | Manual | | Excel | Medium (formulas, charts) | Limited | Computation | | Word | Medium (structured text) | Limited | Manual | | Mixed folder | Very high | Usually none | Manual |
---
The Workspace-State Contract
The authors formalize a contract:
> Each view must be explicitly bound to a specific version of the evolving workspace state.
Four operations must obey it:
1. Parsed Search — version seen during search 2. Native Edit — version operated on during editing 3. Review Diff — versions compared during review 4. Submit Artifact — version finally submitted
Three failure modes are identified:
- Artifact-only: native files preserved but search is hard; agents must page through large files carrying irrelevant context.
- Parsed-only: easy search, but layout, formulas, and visual evidence are lost; edits target an abstract, not the native file.
- Unversioned mutable workspace: agents can overwrite, rename, or delete files, but no persistent diff exists for either model or human review — "What did I change?" becomes unanswerable.
- W_t — Current native workspace files: the authoritative state used for execution and submission.
- C_t — Parsed records, tagged by source path and content hash. If the source hash matches the corresponding file in W_t, the record is marked current; otherwise stale.
- Δ_t = δ(W_0, W_t) — Change diffs between the starting and current workspace, presented in format-specific ways (line diffs for text, row/cell diffs for spreadsheets, slide-level diffs for presentations).
- OfficeQA Pro: numerical QA over US Treasury Bulletins; 104 document-only questions + 29 requiring web evidence; shared workspace of ~697 PDFs (1939–2025); primary metric exact-match Pass@1.
- APEX-Agents: 33 professional worlds (management consulting, investment banking, law); 480 graded tasks; each task folder averages ~166 mixed-format files; primary metrics Pass@1 and mean rubric score.
- Agent design: workspace state should be reported alongside model choice and prompting. Dual parsed/native views are complementary, not competing. Pre-submission review produces a checklist effect that reduces errors.
- Benchmarks: should grade evidence citations, staged edits, final artifact quality, and explicit state transitions from initial to final state.
- Bigger picture: AI is evolving from stateless functions → short-term context → persistent, versioned workspace participants. This raises new requirements for traceability, auditability, reversibility, and authorization on critical operations.
- OfficeQA Pass@1 gain (dual vs. single): 8.3–12.1 pts
- APEX mean rubric gain (dual vs. single): 4.7–9.2 pts
- Gemini 3.1 Pro / OfficeQA SW-Agent: 63.9% (vs. 29.3% baseline)
- GPT-5.4 Nano / APEX SW-Agent: 42.1 (vs. 25.5 baseline)
- Paired review-axis tasks: 57
- OfficeQA workspace size: 697 PDFs
- APEX tasks: 480 across 33 worlds
- APEX avg files per task: 166
- Hua, Y., Na, H., Zhou, Y., Kalose, A., Ayubcha, C., & Lian, L. (2026). *StagedWorkspace: A Versioned Workspace for Knowledge-Work Agents*. 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.
---
StagedWorkspace Architecture
StagedWorkspace maintains three synchronized views:
After each mutation tool batch the system:
1. Hash-scans to identify changed native files 2. Advances W_t to t+1 3. Marks stale entries in C_t (refresh only affected records) 4. Queues asynchronous refresh of the parse cache
Crucially, the parse cache and review diffs are derived views, not independent document copies. The system guarantees a version-aware read-after-write contract: after an edit, native operations target the updated workspace parse, and parsed results either match that version or are explicitly marked stale until refreshed.
---
Experimental Results
Two benchmarks were used:
Full-system performance
| Model | Benchmark | SW-Agent | Published baseline | Gain | |---|---|---|---|---| | Gemini 3.1 Pro | OfficeQA | 63.9% | 29.3% | +34.6 | | GPT-5.4 Nano | APEX | 42.1 | 25.5 | +16.6 |
Crucially, these are same-model comparisons — better workspace management, not bigger models.
Ablations
Read-axis ablation (dual vs. single views):
| Condition | OfficeQA Pass@1 | APEX Mean Rubric | |---|---|---| | Dual view | Highest | Highest | | Artifact-only | −8.3 to −12.1 pts | −4.7 to −9.2 pts | | Parsed-only | Intermediate | Intermediate |
Review-axis ablation (paired across 57 file-edit tasks): visible diffs yielded higher scores when agents could inspect workspace_diff and workspace_file_diff before submitting.
---
Why Version Control Matters for Knowledge Work
1. Cognitive offloading — Miller's Law caps working memory at 7±2 chunks. Git already offloads programmer cognition; StagedWorkspace aims to do the same for non-coding knowledge work.
2. Auditability — humans must understand *what* the agent changed, *on what evidence*, and *why*. Journaled review diffs provide this.
3. Reversibility — mistakes are inevitable; a versioned workspace enables rollback, like git revert.
---
Implications
---
Key Numbers Recap
---
References
*Interpreted 2026-08-20. Feynman-style deep dive.*