StagedWorkspace: A Versioned Workspace for Knowledge-Work Agents

Yining Hua; Hongbin Na; Yifan Zhou; Akshay Kalose; Cyrus Ayubcha; Levi Lian

cs.AIAug 18, 2026

StagedWorkspace: A Versioned Workspace for Knowledge-Work Agents

Authors: Yining Hua, Hongbin Na, Yifan Zhou, Akshay Kalose, Cyrus Ayubcha, Levi Lian

Organizations: Harvard University · Raycaster AI · University of Technology Sydney · University of Washington ·

Abstract

AI agents increasingly perform knowledge work (i.e., produce and modify persistent digital artifacts such as code repositories, documents, spreadsheets, slides, reports), yet the parsed views they search, the native files they edit, the changes they review, and the artifacts they submit can refer to different versions of the same work product. We formulate this as a workspace-state contract: every view should be explicitly tied to a version of the evolving workspace state. Coding agents partly address this need through repository contracts for search, diffs, and tests, whereas an analogous contract is less explicit for PDFs, spreadsheets, slides, notebooks, and mixed-format project folders. We propose StagedWorkspace, a versioned workspace for knowledge-work agents. The workspace binds parsed records and review diffs to content hashes of the native files as they change. In fixed-harness ablations on OfficeQA Pro and APEX-Agents, dual parsed/native access has the highest point estimate for every tested model; relative to the more limiting single view, it improves OfficeQA Pass@1 by 8.3-12.1 points and APEX mean rubric score by 4.7-9.2 points. SW-AGENT scores 63.9% with Gemini 3.1 Pro on OfficeQA and 42.1 with GPT-5.4 Nano on APEX, compared with published same-model scores of 29.3% and 25.5, respectively. A paired review-axis ablation on 57 file-editing tasks further finds higher observed scores when diffs are visible. These results identify workspace state as an experimental variable in knowledge-work agents and motivate benchmarks that score evidence, staged edits, and submitted artifacts as explicit state transitions.

Figures & tables

Figure 1: An explicit workspace-state contract ties native artifacts, parsed views, review state, and submission to the evolving workspace.

Figure 2: Workspace synchronization across three agent-facing views. (a) After editing budget.xlsx at Sheet1!B12 and scanning content hashes, the native workspace contains $W_{t+1}$ , the review view compares $W_{0}$ with $W_{t+1}$ , and a parsed read may still return the prior value with freshness = stale. (b) The parsed result remains stale while $h_{\mathrm{before}}\neq h_{\mathrm{after}}$ and becomes current after asynchronous re-parsing. (c) Matching hashes reuse cached records; in this example, only the mismatched budget.xlsx is re-parsed.

Arm	Originals	Parsed search	Hash sync	Mechanism isolated
Dual	Yes ( $W_{t}$ )	Yes ( $C_{t}$ )	Yes	Synchronized native and parsed access.
Artifact-only	Yes ( $W_{t}$ )	No	N/A	Native access without indexed parsed search.
Parsed-only	No	Yes ( $C_{t}$ )	Yes	Parsed search without native inspection or execution.

Table 1: Read-axis artifact-view arms. Original artifacts are the hydrated workspace source files

W_{t}

; the parsed index is the hash-keyed searchable cache

C_{t}

, with each record labeled current or stale according to whether its source hash matches

W_{t}

Table 2: Complete-system comparisons. In Panel (a), “Full corpus + dual views” means SW-Agent runs on the full OfficeQA corpus with native files and parsed search; Oracle rows receive gold sources. Panel (b) includes matched APEX rows and an Opus reference. N/A = not reported; – = not applicable.

Table 3: Mechanism ablations. Panel (a) reports the OfficeQA Pro read-axis ablation using Pass@1. Panel (b) reports the APEX-Agents read-axis ablation using mean rubric score. Panel (c) reports the APEX-Agents review-axis ablation on paired file-editing tasks. In Panels (a) and (b), lift columns show signed paired lift from dual with the bootstrap half-width in parentheses.

Panel of Figure 3 — Figure 3: Diagnostic cost–score ablations. Lower and further right is better: a run scores more while spending less. Each arrow points from an ablated access configuration to the dual-access result for the same model. Dual access generally improves diagnostic score, although the associated cost change varies across models and benchmarks.

Figure 4: APEX-Agents HarFeast trace. The parsed cache identifies the relevant survey-column guide, and native workbook execution applies that interpretation to the same 3,000-row spreadsheet that is later graded. The example illustrates the APEX mechanism: evidence localization and native execution must refer to the same workspace state.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Tool	Dual	Artifact-only	Parsed-only
view(text)	Parsed pages with freshness metadata	pdftotext /Office fallback	Parsed pages only
view(visual)	Native render	Native render	Unavailable
grep / file_search	Parsed page locators	Raw paths only; weak for PDFs	Parsed page locators
vector_search	Page-chunk embeddings	Disabled	Page-chunk embeddings
python / bash	Native bytes and deliverables	Native bytes and deliverables	No native deliverables

Appendix

Table A.1 : Read-path tool routing by artifact-view arm.

Model	Pass@1 (%)	Pass ×3 (%)	Pass@any (%)
GPT-5.4	37.7	29.0	47.3
GPT-5.4 Mini	27.7	16.5	39.6
GPT-5.4 Nano	25.0	14.6	35.9

Appendix

Table A.2 : APEX-Agents : Historical multi-attempt diagnostic for GPT-family dual runs. Pass@1 is the mean per-attempt pass rate in that diagnostic. Pass ×3 : all three attempts pass. Pass@any : at least one passing attempt. This diagnostic is not used for the main paired ablations.

Bench.	Model	Contrast	Dual wins	Restr. wins	$p_{\mathrm{disc}}$
OfficeQA	GPT-5.4	vs artifact-only	20	6	0.009
OfficeQA	GPT-5.4	vs parsed-only	16	10	0.327
OfficeQA	Gemini 3 Flash	vs artifact-only	17	5	0.017
OfficeQA	Gemini 3 Flash	vs parsed-only	9	5	0.424
OfficeQA	Gemini 3.1 Pro	vs artifact-only	20	9	0.061
OfficeQA	Gemini 3.1 Pro	vs parsed-only	19	7	0.029

Appendix

Table A.3 : Read-axis paired-discordance checks.

p_{\mathrm{disc}}

is McNemar for OfficeQA and a sign test for APEX; these count paired wins and losses as a discrete check on the bootstrap contrasts.

Model	Configuration	Accuracy	Latency (min)	Cost ($)	Tools
GPT-5.4	Artifact-only	52.6 $\,\pm\,$ 8.7	8.1	0.85	34.7
GPT-5.4	Parsed-only (Reducto)	58.6 $\,\pm\,$ 8.3	17.3	0.76	34.3
GPT-5.4	Parsed-only (Databricks)	60.2 $\,\pm\,$ 8.3	7.9	0.94	30.6
GPT-5.4	Dual	64.7 $\,\pm\,$ 8.3	6.1	0.84	30.4
Gemini 3 Flash	Artifact-only	47.7 $\,\pm\,$ 8.7	16.9	0.53	41.7
Gemini 3 Flash	Parsed-only (Reducto)	53.8 $\,\pm\,$ 8.5	18.8	0.63	50.0

Appendix

Table A.4 : OfficeQA Pro : artifact-view ablation efficiency. Rows are grouped by model; latency excludes sandbox startup.

Model	Arm	Latency (min)	Cost ($)	Tools
GPT-5.4	Dual	5.0	1.41	28.1
GPT-5.4 Mini	Artifact-only	8.8	0.33	44.0
GPT-5.4 Mini	Parsed-only	10.3	0.38	45.5
GPT-5.4 Mini	Dual	7.7	0.37	42.2
GPT-5.4 Nano	Artifact-only	11.7	0.13	44.3
GPT-5.4 Nano	Parsed-only	10.6	0.14	36.1

Appendix

Table A.5 : APEX-Agents : artifact-view efficiency diagnostics (latency, agent-run cost, and tool counts per arm). Rubric scores are not repeated here; full-pool scores are in Table 2 and paired ablation scores in Table 3(b) .

Strict: dual $>$ parsed-only and dual $>$ artifact-only ( $n{=}147$ instances)
Mechanism	Count (% of slice)
Workspace file discovery (losing arm missed source)	16/147 (11%)
Within-file navigation (sheet/section/clause)	35/147 (24%)
Table or cell extraction from source	25/147 (17%)
Layout or visual read (PDF/deck structure)	14/147 (10%)
Model propagation or calculation chain	28/147 (19%)

Appendix

Table A.6 : APEX-Agents : documented mechanisms on dual-win instances. Percentages are within each slice.

Domain	Evaluated	All-arm zero	Rate
All domains	452	188	41.6%
Investment Banking	132	82	62.1%
Management Consulting	160	55	34.4%
Law	160	51	31.9%

Appendix

Table A.7 : All-arm-zero tasks by APEX-Agents domain (

452

evaluated tasks). All-arm zero = dual, artifact-only, and parsed-only all score 0 on the same task for at least one completed model triplet.

LLM Agents

Large language model agents use language models to plan, use tools, and act across multi-step tasks and environments.

See all 705 papers

Workspace-Bench 1.0: Benchmarking AI Agents on Workspace Tasks with Large-Scale File Dependencies
May 5, 2026Zirui Tang, Xuanhe Zhou, Yumou Liu +19Computer-Use Agent Benchmarks AI Agent Evaluation
1Shanghai Jiao Tong University · 2ByteDance · 5Independent Researher +2
WorkWorlds: An Infrastructure for Evaluating AI Agents on Workplace Tasks
Sep 20, 2026Yining Hua, Levi LianBenchmark Design AI Agent Benchmarks
Harvard University · Agent Evaluation Science Inc. · Raycaster +1
RunningTab: Direct Workspace Interaction with Environment-Side Tabs
Oct 7, 2026Jinheon Baek, Soyeong Jeong, Yumin Choi +2LLM Agent Reliability LLM Agent Harnesses
KAIST

May 5, 2026J/K move · Enter open · S save

Workspace-Bench 1.0: Benchmarking AI Agents on Workspace Tasks with Large-Scale File Dependencies

Zirui Tang, Xuanhe Zhou, Yumou Liu, Linchun Li+18

1Shanghai Jiao Tong University · 2ByteDance · 5Independent Researher+2

Computer-Use Agent Benchmarks AI Agent Evaluation AI Agent Benchmarks

Open paper AskPDF arXiv

Workspace learning requires AI agents to identify, reason over, exploit, and update explicit and implicit dependencies among heterogeneous files in a worker's workspace, enabling them to complete both routine and advanced tasks effectively. Despite its importance, existing relevant benchmarks largely evaluate agents on pre-specified or synthesized files with limited real-world dependencies, leaving workspace-level evaluation underexplored. To this end, we introduce Workspace-Bench, a benchmark for evaluating AI agents on Workspace Learning involving Large-Scale File Dependencies. We construct realistic workspaces with 5 worker profiles, 74 file types, 20,476 files (up to 20GB) and curate 388 tasks, each with its own file dependency graph, evaluated across 7,399 total rubrics that require cross-file retrieval, contextual reasoning, and adaptive decision-making. We further provide Workspace-Bench-Lite, a 100-task subset that preserves the benchmark distribution while reducing evaluation costs by about 70%. We evaluate 4 popular agent harnesses and 7 foundation models. Experimental results show that current agents remain far from reliable workspace learning, where the best reaches only about 60%, substantially below the human result of 80.7%, and the average performance across agents is only 43.3%.

Abstract

LLM Agents

Explore similar work

Workspace-Bench 1.0: Benchmarking AI Agents on Workspace Tasks with Large-Scale File Dependencies

WorkWorlds: An Infrastructure for Evaluating AI Agents on Workplace Tasks

RunningTab: Direct Workspace Interaction with Environment-Side Tabs