CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork Agents
Authors: Zhongwei Xie, Yiming Du, Jiaxin Bai, Baixuan Xu, Chunyang Li, Xiyu Ren, Wei Fan, Tianshi Zheng, Weizhi Du, Jingqiao Yu, Xingsu Li, Zhaowei Wang, Yangqiu Song
Organizations: The Hong Kong University of Science and Technology · The Chinese University of Hong Kong · Hong Kong Baptist University · Independent Researcher
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.
Figures & tables
Figure 1: Task construction and the teacher-demonstration path. Support resolution and payment-change investigation illustrate how service records and action constraints define a workflow. Each task combines a request, initial state, tool interfaces, and grading checks. Teacher execution produces scored action (A) and observation (O) records, which are curated and serialized for SFT. The same task components support student interaction and online reward computation through the RL interface described in Section 3.4 .
Figure 2: Service coverage and task composition of the construction libraries. (a) Qualitative service and input coverage across both inventories, grouped by function. (b–c) Task families (inner rings) and subtypes (outer rings) for Claw-Eval (602 task identifiers) and ClawMark (263), with shared colors and labels for the largest sectors. The 236 shared identifiers are counted once in each inventory; totals are not additive. Counts describe construction-library snapshots, not benchmark test sizes or checkpoint training-set sizes.
Model
General [-1pt] n=161
Multi-turn [-1pt] n=38
Overall [-1pt] n=199
Score
Pass@3
Pass 3
Score
Pass@3
Pass 3
Score
Pass@3
Pass 3
Claude Opus 4.6
80.6
80.8
70.8
79.6
89.5
68.4
80.4
82.4
70.4
Claude Sonnet 4.6
81.3
81.4
68.3
81.9
89.5
65.8
81.4
82.9
67.8
GPT 5.4
78.3
75.8
60.2
79.0
89.5
60.5
78.4
78.4
60.3
Gemini 3.1 Pro
76.6
80.8
55.9
80.2
92.1
65.8
77.3
82.9
57.8
MiMo V2 Pro
76.0
72.7
57.1
81.0
92.1
60.5
77.0
76.4
57.8
Table 1: Claw-Eval performance on General and Multi-turn tasks. Score, Pass@3, and Pass 3 use a 0–100 scale. Pass@3 and Pass 3 require at least one and all three trials to score ≥0.75 , respectively. Overall combines the two groups with equal weight per task. Published references reproduce Table 3 of Ye et al. [2026] . Parentheses show absolute changes from the same-size base model. Scores and changes are rounded separately to one decimal place. Darker cells indicate higher scores within each color group. Evaluation setup: Section 4.1 .
Model
Tasks
Score ↑
GPT-5.4
100
56.7
Claude Sonnet 4.6
100
53.1
Qwen3.6 Plus
100
48.7
Gemini 3.1 Pro Preview
100
38.9
MiniMax M2.7
100
33.0
Kimi K2.5
100
56.0
Table 2: ClawMark mean task scores (0–100). Website references use OpenClaw Run 1 over 100 tasks [ Evolvent AI, 2026b ] . Parentheses show absolute changes from Qwen3-VL-8B-Instruct. Protocol details appear in Appendix F .
Training condition
General [-1pt] n=161
Multi-turn [-1pt] n=38
Overall [-1pt] n=199
Multimodal [-1pt] n=101
All tasks [-1pt] n=300
Qwen3-VL-8B-Instruct
46.7
45.9
46.6
20.7
37.9
GRPO
52.8 -1pt
72.5 -1pt
56.6 -1pt
22.1 -1pt
45.0 -1pt
GDPO
53.0 -1pt
74.9 -1pt
57.2 -1pt
21.0 -1pt
45.0 -1pt
Dr.GRPO
51.2 -1pt
67.1 -1pt
54.3 -1pt
21.0 -1pt
43.1 -1pt
GiGPO
59.7 -1pt
64.8 -1pt
60.7 -1pt
21.0 -1pt
47.3 -1pt
Table 3: Claw-Eval results after supervised warm-up and RL. Scores on the 300-task benchmark use a 0–100 scale. Overall combines General and Multi-turn; All tasks also includes Multimodal. The General, Multi-turn, and Overall base scores are shared with Table 1 . Parentheses show absolute changes from Base for the warm-up-plus-RL pipeline. Scores and changes are rounded separately to one decimal place. Bold indicates the highest score in each column.
SFT data
General [-1pt] n=161
Multi-turn [-1pt] n=38
Overall [-1pt] n=199
Base (no SFT)
46.7
45.9
46.6
20%
50.1 -1pt
53.6 -1pt
50.8 -1pt
50%
53.4 -1pt
64.9 -1pt
55.6 -1pt
100% (full)
60.9 -1pt
77.8 -1pt
64.1 -1pt
Table 4: Claw-Eval scores (0–100) for Qwen3-VL-8B-Instruct with different SFT data fractions. The specified fraction is randomly sampled within each SFT stage; parentheses show gains over Base. Base and full-data scores are the three-trial results from Table 1 . Overall weights tasks equally. Scores and gains are rounded separately to one decimal place.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Comparison of RL training methods on our constructed tasks. Rows show GRPO, GDPO, GiGPO, and Dr.GRPO; columns show training reward, policy loss, and response length over approximately three epochs. The horizontal axis counts cumulative processed training samples, including repeated sampling of a prompt across epochs. Curves are smoothed using a moving average over 8 logged steps.
Component
Generated tool workflows
Dialogue demonstrations
Task source
Coverage plans, generated specifications, and reusable services
Independently constructed dialogue scenarios and user personas
Interaction
An initial request followed by agent tool calls and service observations
A persona-based user simulator exchanges turns with the agent
Grading
Deterministic checks of tool calls, arguments, and audit records
An LLM judge evaluates clarification and final-answer quality
Preparation
Export execution records and convert them to student messages
Extract recorded dialogue trials, select demonstrations, and normalize messages
Appendix
Table 5: Preparation paths for tool-workflow and dialogue demonstrations collected on our constructed training tasks.
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.
Scaling interactive environments is important for training long-horizon agents through reinforcement learning. These environments often require further testing and refinement after initial generation, and correcting defects in a task may also require updating its work context and verifier. However, existing construction pipelines offer limited support for such iterative refinement while keeping these components aligned. We introduce WorkRover, a framework for training long-horizon cowork agents through recursive environment evolution. Its harness-driven synthesis enables construction agents to jointly build and validate semantically aligned work contexts, executable tasks, and evidence-grounded verifiers, adapting their actions to workspace state and execution feedback rather than following a fixed construction workflow. To support environment evolution, we introduce Factorial Compatibility Replay (FCR), which uses rollout feedback to evaluate combinations of component revisions through dependency-closed replay and selects the smallest validated repair among tested candidates while preserving task requirements and evaluation constraints. To provide finer-grained training signals, we propose Verifier-to-Turn Credit Assignment (VTCA), which augments trajectory-level GRPO advantages with signals from deterministic workspace-progress checks and assigns a bounded share of positive-progress credit to prerequisite turns using recorded tool-output and artifact dependencies. Using WorkRover, we post-train Qwen3.5-35B-A3B to obtain WorkRover-35B-A3B and evaluate it on several cowork-agent benchmarks. Its reported scores exceed those of Qwen3.5-35B-A3B by 26.27 percentage points on JobBench, 29.32 percentage points on OfficeQA Pro, and 413.25 Elo points on GDPval v2. These results suggest that WorkRover offers a promising foundation for scalable training of long-horizon cowork agents.
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.