PaperlayerView exact revision on arXiv

Source-linked AI summary

Cowork-Gym: Synthesizing Interactive Work Environments at Scale

Anni Zou, Cheng Chen, Jiaji Deng, Weikang Zhou, Xinji Mai, Yue Wu, Jianing Lin, Lipeng Xie, Lixiong Qin, Qingyue Zhao, Que Shen, Shuchang Tao, Xuchen Pan, Yaliang Li, Yunpeng Zhai, Yushuo Chen, Zhihai Wang, Zhuo Zhang, Zhaoyang Liu, Bolin Ding, Jingren Zhou

arXiv:2610.17670v1cs.AI

TL;DR

Cowork environment synthesis is difficult because tasks, changing workspace state, user behavior, and outcome checks must remain consistent. Cowork-Gym builds executable environments with iterative validation and stateful user simulation, then uses them to post-train agents that improve across benchmarks.

  • Problem

    Synthesizing Cowork environments at scale requires consistent task requirements, executable workspaces, user behavior, and outcome checks as workspace state and requirements change.

    Source: §1 · p. 2

  • Method

    Cowork-Gym combines taxonomy-guided and resource-grounded task generation, build–execute–revise validation, and a stateful user simulator that changes dialogue and workspace state with agent progress.

    Source: §1 · p. 2, §8 · p. 15

  • Results

    All three Qwen model scales improve on average across six agentic benchmarks; Ours-27B ranks second overall and leads PawBench and ClawEval, surpassing models with over 100B total parameters.

    Source: §1 · p. 2, §8 · p. 15

  • Takeaways & Limitations

    Scalable executable environments can provide a data-driven path toward self-evolving agents that adapt as tools, workspaces, and user requirements change.

    Source: §8 · p. 15

    The synthetic pool is not representative coverage: it emphasizes browser, file, data, and office workflows, while complex visual interfaces and richer audio or video tasks are underrepresented.

    Source: §7 · p. 15

Abstract

from arXiv · show

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.

1 INTRODUCTION

Cowork work unfolds across tools and a persistent workspace while users can change requirements and inputs, making consistent, executable environments costly to prepare and check. Cowork-Gym synthesizes environments with iterative validation and a stateful user simulator; models trained on the resulting data improve broadly across agentic benchmarks.

1 INTRODUCTION

  • Scaling synthesis is difficult because task requirements, executable workspaces, user behavior, and outcome checks must stay consistent as interactions change.A final response may appear correct even when a required artifact is missing or the workspace is left in an incorrect state.Source: §1 · p. 2
  • Cowork-Gym combines taxonomy-guided and resource-grounded task generation with a build–execute–revise loop that checks and repairs executable environments.Sandbox execution and verifier checks identify inconsistencies before trajectory collection.Source: §1 · p. 2
  • The stateful user simulator adapts dialogue and shared-workspace edits to agent progress, requiring agents to respond to changing requirements and inputs.For example, the simulator may replace a source table after reviewing an initial analysis.Source: §1 · p. 2
  • 17 of 18 matched model–benchmark comparisons improve over base checkpoints; Ours-27B ranks second overall and leads PawBench and ClawEval.On both benchmarks, Ours-27B surpasses evaluated models with over 100B total parameters.Source: §1 · p. 2
  • The framework synthesizes 64.9K environments across 14 top-level scenarios spanning office, software, data, and other personal and professional workflows.Collected trajectories and verification signals are used to post-train Qwen models with supervised fine-tuning, reinforcement learning, and optional on-policy distillation.Source: §1 · p. 2

2 PROBLEM FORMULATION

A Cowork task is specified as an environment, realized through an agent-and-harness rollout, and converted into a training sample through filtering and transformation. The environment packages the task and its evolving interaction context, while trajectories record the resulting workspace activity.

2 PROBLEM FORMULATION

  • Cowork data has three levels: an environment specifies the task, a model–harness configuration produces a trajectory, and filtering and transformation yield a training sample.The same environment may produce different trajectories under different policies or harnesses.Source: §2 · p. 3
  • An environment couples the task, initial workspace, interfaces, stateful user policy, and hybrid verifier, and is packaged with runtime, reference solution, and user profiles.The verifier combines deterministic artifact checks with LLM-based evaluation of semantic correctness and task-relevant process evidence.Source: §2 · p. 3, Figure 2 · p. 3
  • Pairing an environment with an agent policy and harness produces a trajectory recording workspace changes, agent actions, user interventions, tool calls, and tool results.The harness mediates observations, tool schemas, execution, context management, and stopping behavior.Source: §2 · p. 3
  • Before training, a filtering operator checks task outcomes and structural validity; retained trajectories are normalized into the downstream training representation.The synthesis overview places filtering after user-simulated trajectory generation and before SFT, RL, and OPD.Source: §2 · p. 3, Figure 3 · p. 4

3 SCALABLE COWORK DATA SYNTHESIS

This section traces how Cowork-Gym turns task specifications into validated environments, interactive trajectories, and training samples.

3.1 TASK SPECIFICATION

  • Taxonomy-guided and resource-grounded routes create task specifications, respectively controlling scenario coverage or deriving requirements from existing resources and their dependencies.The taxonomy spans six dimensions; resource-grounded tasks are retained when their source material contributes materially to completion.Source: §3.1 · pp. 4–5

3.2 BUILD–EXECUTE–REVISE LOOP

  • The build–execute–revise loop constructs the environment, tests it for defects, and routes each defect to its responsible constructor for repair until release criteria pass or the iteration budget is exhausted.Specialized Builder, Verifier Constructor, and Issue Detector agents operate with separately configured prompts, permissions, and sandboxes.Source: §3.2 · p. 5

3.2.1 BUILDER

  • The Builder creates the initial workspace, task interfaces, and private user policy while keeping privileged state, verifier implementation, and construction metadata outside the acting agent’s sandbox.It records dependencies so user changes can invalidate and trigger updates to dependent artifacts.Source: §3.2.1 · p. 5

3.2.2 VERIFIER CONSTRUCTOR

  • The Verifier Constructor independently maps task requirements to weighted checks, combining deterministic workspace and trajectory tests with semantic and procedural evaluation.The hybrid verifier is self-validated on a reference completion and controlled failure cases before review.Source: §3.2.2 · pp. 5–6

3.2.3 ISSUE DETECTOR

  • The Issue Detector tests environments through execution and rule audits, then routes defects for repair and rejects candidates that remain unresolved after the iteration budget.Execution review checks verifier behavior on valid, incomplete, malformed, and untouched states; rule audits check package, permission, and leakage criteria.Source: §3.2.3 · p. 6

3.3 STATEFUL USER SIMULATOR

  • The stateful user simulator uses private user context to respond to agent progress with dialogue and authorized workspace edits, which can invalidate earlier observations and require dependent artifacts to be revised.The same initial query can yield different follow-ups because interventions depend on actual progress.Source: §3.3 · pp. 6–7

3.4 TRAJECTORY GENERATION AND FILTERING

  • Rollouts keep each environment fixed while varying the acting policy or harness, and filtering checks record integrity, verified task outcomes, and whether success claims match observed results.Intermediate tool errors are allowed when the agent recovers and reaches a verified outcome.Source: §3.4 · p. 7
  • Model-conditioned selection prioritizes trajectories for a target student using NLL or RSR under a fixed retention budget, with length-binned retention limiting shifts in trajectory-length distributions.NLL retains the central fraction in each bin, whereas RSR retains trajectories with the lowest scores.Source: §3.4 · p. 8

4 AGENTIC POST-TRAINING

This section describes direct post-training with supervised fine-tuning and outcome-based reinforcement learning, alongside on-policy distillation. The distillation route trains on student-generated interaction histories using a same-size teacher.

Supervised fine-tuning

  • SFT trains on filtered trajectories, predicting agent-generated messages and tool calls while using public user messages, tool observations, and workspace events as context.The training mixture includes successful executions, revision-driven interactions, and grounded recoveries from tool errors.Source: §4 · p. 8

Outcome-based reinforcement learning

  • Outcome-based RL applies GRPO to SFT checkpoints using a weighted geometric mean of executable-verification and semantic-correctness scores, assigning zero reward if either score is zero.The geometric mean penalizes imbalanced score pairs more strongly than an arithmetic mean with the same normalized weights.Source: §4 · p. 8

On-policy distillation

  • On-policy distillation trains a base-initialized student to match a same-size SFT or SFT+RL teacher’s next-token distributions on agent-generated tokens from student-collected trajectories.User messages, tool observations, and workspace events provide context but are not prediction targets.Source: §4 · p. 9

5 EXPERIMENTS

The experiments assess synthesized-task coverage, training gains across six agentic benchmarks, and how training choices affect transfer and capability retention. The setup uses six benchmarks and compares three Qwen model scales.

5.2 DATA SCALE AND DIVERSITY

  • The pool contains 64,888 environments across 14 scenarios and 17 capabilities; filtering retains 17,312 of 77,729 raw trajectories, with a 10-step median in the evidence-valid set.Coverage is non-uniform, with office productivity and software engineering the largest scenario groups. Multi-turn trajectories have a 15-step median and 32.69% exceed 20 tool-using steps, versus 10 steps and 15.80% for single-turn trajectories.Source: §5.2 · pp. 9, 11, Table 1 · p. 10

5.3 MAIN RESULTS

  • Across six benchmarks, 9B, 27B, and 35B-A3B improve over matched baselines by +8.7, +6.4, and +4.1 points, respectively, with gains in 17/18 comparisons; Ours-27B ranks second overall and leads PawBench and ClawEval.Ours-27B’s six-benchmark mean is 0.6119 versus 0.6157 for GLM-5.2, and it leads both benchmarks against evaluated models with over 100B total parameters.Source: §5.3 · p. 11, Table 2 · p. 11

5.4.1 SUPERVISED ENVIRONMENT EXPOSURE

  • On ClawEval, all three models improve across the three metrics during supervised training; 27B finishes highest on all three, while 35B-A3B’s pass3 continues improving after other metrics plateau.The 9B model has the largest Score (+4.66) and pass@3 (+5.00 percentage points) gains; 35B-A3B has the largest pass3 endpoint gain (+5.80 percentage points).Source: Figure 6 · p. 12, §5.4.1 · p. 12

5.4.2 INTERACTIVE SUPERVISION

  • Adding multi-turn trajectories to single-turn training improves average scores on all three benchmarks, with the largest gain on ClawMark (+3.23 points).Source: Table 3 · p. 12, §5.4.2 · p. 12

5.4.3 CROSS-HARNESS TRANSFER

  • Training on either collection harness improves pass rates on all three evaluation harnesses, including unseen Hermes; combining sources yields 31.8% pass rates on both QwenPaw and OpenClaw.OpenClaw-only training gains +9.0 points on Hermes and +10.0 on QwenPaw; QwenPaw-only training gains +7.2 on Hermes and +9.1 on QwenPaw.Source: Table 4 · p. 12, §5.4.3 · p. 13

5.4.4 TRAJECTORY SELECTION

  • At 70% retention, NLL selection improves every reported metric across all three model sizes, while the strongest criterion shifts from NLL at 9B and 27B to RSR on most 35B-A3B metrics.Selection within length bins makes the effect less readily attributable to length bias.Source: §5.4.4 · p. 13, Table 5 · p. 13

5.5 ON-POLICY DISTILLATION

  • On ClawEval, OPD improves pass3 by +4.5 points over base, recovering about 80% of the teacher’s gain; on GPQA-Diamond, it stays near base and exceeds the SFT teacher by +4.54 points.Source: §5.5 · p. 14, Figure 7 · p. 14

6 RELATED WORK

Related work covers executable agent environments, cowork evaluation, stateful user simulation, and agent post-training. Cowork-Gym combines environment construction, user policies, harness-aware rollouts, and verified training signals.

Executable agent environments

  • Executable testbeds and environment-synthesis systems target web, mobile, and desktop agents; Cowork-Gym applies these principles to personal work by jointly constructing workspaces, interfaces, user policies, and verifiers.Source: §6 · p. 14

Workspace and cowork evaluation

  • Benchmarks cover file collections, professional workflows, long-horizon tasks, evolving service state, and trajectory or process evidence; this work instead synthesizes training environments and verified interaction data.Source: §6 · p. 14

Stateful user–agent interaction

  • Prior user simulators combine tool execution with multi-turn interaction, while human-grounded studies report excessive cooperation and limited behavioral diversity; Cowork-Gym models task-conditioned interventions and workspace changes.Source: §6 · pp. 14–15

Agent post-training and harness effects

  • Interactive post-training spans environment-grounded RL and on-policy distillation, while harness benchmarks show model–harness variation; Cowork-Gym treats the harness as a rollout variable and derives training signals from validated environments and filtered trajectories.Source: §6 · p. 15

7 LIMITATIONS AND FUTURE DIRECTIONS

Cowork-Gym’s synthesized task pool has important coverage boundaries, and its verifier’s reliability limits data selection and reinforcement learning.

Task coverage

  • Synthetic tasks may omit cultures, languages, accessibility needs, and failure modes, while complex visual interfaces and specialized desktop, audio, and video workflows remain underrepresented.The pool emphasizes browser, file, data, and office workflows; its scale should not be treated as representative coverage.Source: §7 · p. 15

Verifier reliability

  • Verifier quality constrains both data selection and reinforcement learning; programmatic checks can be brittle and language-model graders biased or overly permissive.Positive, negative, and mutation tests reduce but do not eliminate these risks, particularly for open-ended artifacts.Source: §7 · p. 15

8 CONCLUSION

Cowork-Gym synthesizes executable environments and filtered trajectories for post-training, with a stateful simulator supporting dialogue and workspace edits. The resulting Qwen models improve across six benchmarks, while scalable environment construction is presented as a path toward adapting to changing work settings.

8 CONCLUSION

  • Cowork-Gym combines taxonomy-guided and resource-grounded generation with a build–execute–revise loop to construct and validate complete executable environments.The framework synthesizes 64.9K environments across 14 application scenarios and collects filtered trajectories for post-training.Source: §8 · p. 15
  • All three post-trained Qwen sizes improve over their base checkpoints on average across six benchmarks; Ours-27B ranks second overall and leads PawBench and ClawEval, surpassing evaluated models above 100B parameters on both.The models are 9B, 27B, and 35B-A3B; the 27B comparison is specific to PawBench and ClawEval.Source: §8 · p. 15
  • The authors present scalable environment construction as a data-driven path toward agents that adapt as tools, workspaces, and user requirements evolve.The environment pool can expand to supply new tasks, interactions, and verification signals for subsequent training.Source: §8 · p. 15

A TASK TAXONOMY

The taxonomy guides task generation and dataset coverage by separating task operations from application domains and other constraints.

A TASK TAXONOMY

  • A versioned taxonomy controls task generation and dataset coverage.Source: A TASK TAXONOMY · p. 20
  • The taxonomy separates the operations required by a task from its application domain and other task-level constraints.Source: A TASK TAXONOMY · p. 20
  • Taxonomy labels specify the task to construct, not a solution trace or events observed during rollout.Source: A TASK TAXONOMY · p. 20

A.1 TAXONOMY OVERVIEW

The taxonomy has six dimensions, with task attributes generally allowing multiple values while each task receives one primary application path.

A.1 TAXONOMY OVERVIEW

A.2 CAPABILITIES

The appendix separates capability labels from application scenarios: a task can combine operations without requiring a new capability label, while each task receives a path through a scenario hierarchy.

A.2 CAPABILITIES

  • Capability labels are independent of application scenarios, so combining operations within one domain does not require a new domain-specific capability label.Source: §A.2 · p. 20, Table 7 · p. 20

A.3 APPLICATION SCENARIOS

  • Each task receives one primary scenario path through a three-level hierarchy of broad domain, subdomain, and specific use case.Table 8 summarizes the top two levels and gives a representative leaf for each domain.Source: §A.3 · p. 21, Table 8 · p. 21

B DATA SOURCES

The appendix describes the public and versioned sources used for the environment file pool and presents construction and quality-control prompts for the Builder, Verifier Constructor, and Issue Detector.

B DATA SOURCES

C PROMPTS FOR ENVIRONMENT CONSTRUCTION AND QUALITY CONTROL

  • The abridged construction prompts preserve role definitions, workspace rules, user simulation, verifier construction, and quality-control procedures while omitting repeated engineering instructions and deployment-specific cases.Source: C PROMPTS FOR ENVIRONMENT CONSTRUCTION… · p. 21

C.1 BUILDER

  • The Builder creates the visible workspace and runtime setup, prepares user context and collaborative updates, derives a grounded reference answer by executing the task, and restores the intended initial state.Later turns should introduce new information or changed inputs or requirements rather than split deterministic cleanup steps.Source: §C.1 · p. 22

C.2 VERIFIER CONSTRUCTOR

  • The Verifier Constructor derives expected outcomes from task inputs and combines deterministic output checks with trajectory-level criteria, then tests the verifier on correct and controlled failure states.Checks should prioritize semantic correctness and red-line constraints over superficial file existence.Source: §C.2 · pp. 22–23

C.3 ISSUE DETECTOR

  • The Issue Detector actively inspects and reproduces each task, audits whether the verifier separates valid completions from incomplete or shortcut states, and routes evidence-based defects to the responsible constructor.It does not repair the package directly.Source: §C.3 · pp. 23–24

Related papers

Loading 2610.17670v1…