Rapid advances in cowork agents are shifting human-AI collaboration toward real-world automation, delivering software, designs, and other artifacts. However, such productivity work demands both high-quality delivery on complex productivity tasks and continued monitoring and decision-making over long horizons, a combination that remains an open challenge. We introduce CoworkOS, a general-purpose harness for long-horizon work that separates task strategy from execution management: the model decides actions, delegation, and acceptance, while the runtime maintains state and feedback. Its three-layer architecture enables direct execution or delegation to Subagents, synchronously or asynchronously, under stable responsibilities. Through a shared application layer, humans and agents inspect and revise the same evolving documents, presentations, CAD models, and Blender scenes, so human-in-the-loop participation can enter both through conversation and through direct modification of the work itself. To sustain continuity across agents and sessions, CoworkOS further introduces hierarchical workspaces and persistent semantic memory, so working files, key decisions, and long-running Job logs remain shareable, reusable, and resumable after interruption. Across 12 benchmarks and 6 categories of real-world applications, CoworkOS proves highly competitive with frontier agent applications such as Codex and DeepSeek Harness. As a sustained-work case study, CoworkOS uses GLM-5.3 as its backbone to conduct RSI-driven automated post-training on Qwen3.6-35B-A3B. Over five rounds, the model's average score across eight benchmarks rose from 33.19 to 42.95 (+9.76). Looking ahead, we believe CoworkOS opens a meaningful path toward general cowork-agent intelligence.
Figures & tables
Figure 1: Overview of 24 application cases produced with CoworkOS across six categories: frontend websites, games, Office presentations, interactive visualizations, CAD, and Blender rendering. Two views of each case illustrate different pages, scenes, or inspection states. The montage provides a visual overview; harness comparisons and detailed inspection evidence follow in Figure 6 and Appendix A.2 .
Figure 2: Overview of CoworkOS . (a) Two complementary requirements for cowork agents: complex productivity tasks and continued monitoring and decision-making. (b) Human–agent cowork through shared, evolving work. (c) Performance overview of CoworkOS alongside other frontier cowork harnesses across 10 benchmarks in four domains. (d) Recursive self-improvement (RSI) via automated post-training over five improvement rounds. Details are provided in Section 4.3 and Appendix A.3 .
Figure 3: The CoworkOS harness and its shared action space. Main directs and reviews work, while the Coordinator Runtime manages synchronous or asynchronous Subagent execution and feedback. Main and Subagents share tools, application services, and an agent workspace. Asynchronous delegation lets Main continue other work. Application services support native resource editing, revision management, and conflict detection.
Figure 4: Task workspace structure and collaborative editing. (a) Selected file paths and their owners or record managers across agent, Run, and session scopes. (b) Iterative CAD refinement through user instructions, agent updates, and direct human edits to a shared model. (c) Task references and bindings link to independently stored application resources with persistent content, revisions, and edit history.
Figure 5: Blender integration in CoworkOS . The agent conversation (left) records code execution, updates to workspace files, and image inspection. Blender’s native viewport (right) shows the resulting palace model in an editable 3D scene.
Domain
Benchmark
OpenHands
Codex
DeepSeek Harness
CoworkOS
Tool Use
SkillsBench
63.30
53.48
56.75
64.34
τ3 -Bench Banking
40.20
46.39
55.67
49.48
AutomationBench
50.50
30.33
53.50
51.83
Average
51.33
43.40
55.31
55.22
Productivity
JobBench
55.02
23.42
56.15
59.88
Harvey Legal Agent Bench
23.75
12.70
22.30
22.74
Table 1: Benchmark evaluation across twelve tasks in four domains. Each domain ends with an equally weighted Average over the listed benchmarks. GDPVal reports Elo over all 220 scheduled tasks; its tie-adjusted win rate (percent) enters the Productivity average (Appendix A.1 ). Bold and underlining mark the highest and second-highest scores per row, including ties.
Figure 6: Qualitative comparisons across 24 cases: (a) frontend websites, (b) games, (c) Office presentations, (d) interactive visualizations, (e) CAD, and (f) Blender rendering. Each case shows Codex, DeepSeek Harness (DSH), and CoworkOS from left to right. Complete interactive artifacts are available in the online showcase .
Figure 7: Per-benchmark performance across R0–R5 in RSI-native automated post-training. The 2×4 panels show Automation, BFCL v4, OfficeQA, Tau3 Banking, Harvey LAB, SkillsBench, GDPVal, and JobBench. Dashed lines mark the R0 reference for each task. Labels identify the initial and final scores and any intermediate maximum distinct from the final score. The horizontal axis denotes iteration round.
Figure 8: Average score across eight benchmarks and key changes in RSI-native automated post-training. The curve connects observed averages at R0–R5; the dashed line marks the initial seed-SFT baseline. Each annotation integrates the corresponding round’s checkpoint start, epochs, optimization steps, data recipe, and strategy changes. Sampling multipliers are reported separately from epochs. R0 and R1 start independently from the base model; R2–R5 continue from the preceding checkpoint. These changes describe concurrent interventions rather than isolated causal effects; round index does not represent elapsed time. Detailed training records appear in Appendix A.3 .
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9: Frontend website interaction evidence. Columns show Codex, DSH, and CoworkOS . (a) Kitchen inspection with room information. (b) Exploded pack views, followed by feedback from the recorded item-insertion sequence. Each panel pairs a scene view with a native interface crop; English notes below the crops summarize the observations.
Figure 10: Game scenes and interaction details. Columns show Codex, DSH, and CoworkOS . (a) Interiors and flight routes in Paper Glider . (b) Slicing effects and environments in Fruit Slicer , including a frame from pointer-driven play. Native views are scaled proportionally without retouching.
Figure 11: Office content-page details. Columns show Codex, DeepSeek Harness (DSH), and CoworkOS . (a) Mogao Caves metric groups, from pages 2 / 3 / 2. (b) Aral Sea cause explanations, from pages 5 / 3 / 5. Rectangular crops retain each complete metric group or relation representation; native text and graphics are unchanged.
Figure 12: Interactive visualization evidence. Columns show Codex, DSH, and CoworkOS . (a) Expanded CREMA Espresso assemblies and component-description crops; DSH uses its native zoom-out control. (b) CSA2 Attention in Reuse mode at Selection, with native explanation crops. English notes summarize the observations; native image content is unchanged.
Figure 13: Additional CAD evidence. Columns show Codex, DSH, and CoworkOS . (a) Peripheral regions of assembled turbofan engines. (b) Robotic-joint housing and output interfaces. Each row uses the same inspection view and crop rule after axis alignment. Authored cutaways, transparency, and opaque surfaces are retained. Unretouched crops demonstrate visible structural detail; the shared half-sections in Figure 6 e expose the internals.
Figure 14: Additional Blender scene evidence. Columns show Codex, DSH, and CoworkOS . (a) Corresponding main-hall facades after native orbit and zoom. (b) Vehicle and street-dressing crops from alternate views, retaining the visible bus–road gap in CoworkOS . Views retain the delivered models and presentation settings; they are not calibrated lighting comparisons. Only rectangular crops and proportional scaling are applied.
Benchmark
R0
R1
R2
R3
R4
R5
Δ R5–R0
Automation
11.00
35.50
30.83
39.49
41.83
44.15
+33.15
BFCL v4
63.48
61.81
61.76
62.11
61.76
70.83
+7.35
OfficeQA
55.64
53.38
52.63
56.14
59.40
62.11
+6.47
τ3 -Banking
14.43
19.59
16.49
18.57
22.68
22.68
+8.25
Harvey LAB
3.33
2.50
6.78
7.53
8.42
11.08
+7.75
SkillsBench
20.85
30.38
23.18
34.16
33.84
35.76
+14.91
Appendix
Table 2: Complete results for AI-automated post-training across R0–R5 and the final change relative to the initial baseline.
Cowork agents navigate dependencies across tools and files while adapting to changing user requirements and edits to shared workspaces. Scaling such environments requires maintaining consistency among task objectives, executable resources, user behavior, and success criteria. We introduce Cowork-Gym, a framework that synthesizes Cowork environments through a build-execute-revise loop that uses execution evidence and review feedback to validate and repair generated environments. To capture realistic Cowork dynamics, a stateful user simulator responds to the agent's actions through dialogue and direct edits to the shared workspace. We synthesize 64.9K environments across 14 application scenarios and use them to post-train 9B, 27B, and 35B-A3B Qwen models with supervised fine-tuning, reinforcement learning, and on-policy distillation. Averaged across six agentic benchmarks, these models outperform their corresponding base checkpoints by 8.7, 6.4, and 4.1 points, respectively. Ours-27B ranks second overall among the evaluated models and achieves the highest average scores on PawBench and ClawEval, outperforming models with over 100B total parameters on both benchmarks.
AI coworkers must coordinate tools, communicate across turns, and act on changing workspace state while respecting workflow constraints. Constructing training experience for these capabilities requires coherent requests, service records, tool interfaces, and grading checks. We present CoworkEnv, a capability-guided framework for synthesizing executable training environments for AI coworkers. Workflow requirements guide task construction, and programmatic validation with feedback repair checks generated specifications and grading behavior. Teacher executions yield curated demonstrations for supervised fine-tuning (SFT), while student interactions provide rollouts and rewards for reinforcement learning (RL). Using these resources, we train compact PocketWorker models. For Qwen3-VL-8B-Instruct, SFT raises the mean score from 46.6 to 64.1 on Claw-Eval, including a 31.9-point gain on Multi-turn tasks. An SFT-only extension to ClawMark demonstrates applicability across coworker benchmarks. We further compare RL training dynamics across configurations that vary reward processing and credit assignment. Improving training rewards provide preliminary evidence that the constructed environments support policy optimization.
Zhongwei Xie, Yiming Du, Jiaxin Bai +10
The Hong Kong University of Science and Technology · The Chinese University of Hong Kong · Hong Kong Baptist University +1
Language-model agents are increasingly used as persistent coworkers that assist users across multiple working days. During such workflows, the surrounding environment may change independently of the agent: new emails arrive, calendar entries shift, knowledge-base records are updated, and evidence appears across images, scanned PDFs, audio, video, and spreadsheets. Existing benchmarks do not adequately evaluate this setting because they typically run within a single static episode and remain largely text-centric. We introduce \bench{}, a benchmark for coworker agents built around multi-turn multi-day tasks, a stateful sandboxed service environment whose state evolves between turns, and rule-based verification. The current release contains 100 tasks across 13 professional scenarios, executed against five stateful sandboxed services (filesystem, email, calendar, knowledge base, spreadsheet) and scored by 1537 deterministic Python checkers over post-execution service state; no LLM-as-judge is invoked during scoring. We benchmark seven frontier agent systems. The strongest model reaches 75.8 weighted score, but the best strict Task Success is only 20.0%, indicating that partial progress is common while complete end-to-end workflow completion remains rare. Turn-level analysis shows that performance drops after the first exogenous environment update, highlighting adaptation to changing state as a key open challenge. We release the benchmark, evaluation harness, and construction pipeline to support reproducible coworker-agent evaluation.