Long-horizon cowork agents

Scaling Long-Horizon Cowork Agent Training
Via Recursive Environment Evolution

Hang He^*Linting Wang^*Guanming Liu^*Bo Mao^*Jinxiu Liu^*Yiran Suo^*Maosen ZhouZhiheng XiGuanting Dong

Zhou Shao^†Rui ZhengJie ZhouChengcheng WanTao Gui^†Qi ZhangXipeng QiuXuanjing Huang

Fudan UniversityEast China Normal UniversityRenmin University of China

Shanghai Qiji Zhifeng Co., Ltd.Shanghai Innovation Institute

^* Equal contribution ^† Corresponding authors

Paper GitHub Results BibTeX

JobBench36.89%+26.27 percentage points

OfficeQA Pro48.87%+29.32 percentage points

GDPval v21218 Elo+413.25 Elo points

WorkRover-35B-A3B · Improvements over the Qwen3.5-35B-A3B backbone

The idea

Better environments. More useful feedback.

Train cowork agents by improving the environments they learn from.

Abstract

Scaling interactive environments has become increasingly 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. Harness-driven synthesis jointly builds and validates semantically aligned work contexts, executable tasks, and evidence-grounded verifiers. Factorial Compatibility Replay (FCR) evaluates combinations of component revisions and selects the smallest validated repair among tested candidates. Verifier-to-Turn Credit Assignment (VTCA) augments trajectory-level GRPO advantages with deterministic workspace-progress signals and assigns a bounded share of positive credit to prerequisite turns.

We post-train Qwen3.5-35B-A3B to obtain WorkRover-35B-A3B, improving over its backbone by 26.27 percentage points on JobBench, 29.32 points on OfficeQA Pro, and 413.25 Elo points on GDPval v2.

Construct

Ground tasks in connected files, professional skills, and executable applications.

Harness-driven synthesis

Evolve

Use execution feedback to repair task environments while keeping their components aligned.

Factorial Compatibility Replay

Learn

Reward checked workspace progress and the earlier turns that supplied required evidence.

Verifier-to-Turn Credit Assignment

30,115Verified SFT trajectories

3,200RL tasks after filtering

317Professional roles

5External benchmarks

How it works

From a workspace to a training environment

One harness coordinates construction, validation, repair, and learning.

WorkRover overview: work context construction, long-horizon task synthesis, verifier synthesis and validation, and recursive environment evolution. — **Harness-driven construction.** Work Contexts, Tasks, and Verifiers are built together, with recorded dependencies and independent validation.

FCR · Environment evolution

Repair the components together

A task failure can reveal a mismatch across its files, instructions, execution harness, or verifier. FCR freezes repair proposals, replays their combinations with dependency-induced updates, and checks the resulting environments.

Preserve task requirements and evaluation constraints.
Select the smallest validated repair among tested candidates.
Recheck affected tasks for regressions.

Matched repair budget caps

Natural defects · Safe resolution rate

Local-only

42.0%

FCR
73.0%

81.2%Controlled-defect SRR

0.7%Cross-task regression

Rates from the maintenance evaluation in Table 2. Model-token cost is higher for FCR than for local-only repair.

FCR analysis of joint repair necessity, outcomes on natural defects, stability across three generations, and rollout success over five repair iterations. — **Repeated environment evolution.** Joint revisions, sealed audits, and fresh rollouts evaluate the reliability of each repair.

VTCA · Turn-level learning

Give credit to checked progress

Frozen deterministic checks compare workspace states before and after each turn. Positive progress stays primarily with the turn that produced it, with a bounded share assigned to earlier turns whose exact tool outputs or artifact versions were used. Negative increments stay local. These signals shape GRPO advantages without intermediate judge calls.

VTCA training pipeline: same-task rollout sampling, terminal evaluation, workspace progress checks, prerequisite attribution, and turn-balanced policy updates. — **Verifier-to-Turn Credit Assignment.** At least half of each positive progress increment remains with the producing turn; only recorded direct dependencies qualify for prerequisite credit.

Experiments

Transfer across five cowork benchmarks

Occupational tasks, professional deliverables, app workflows, and document reasoning.

JobBench

Score (%) · higher is better

Reported model scores

All scores (CSV)

Model

Reading these results. Scores reproduce Table 1 of the manuscript. Published reference scores retain their source-specific task subsets, evaluators, and harnesses. GDPval v2 is measured in Elo; the other four benchmarks are percentages. A dash indicates an unreported score.

What does turn-level feedback add?

With the SFT initialization, task and seed schedules, and rollout budget fixed, VTCA improves over episode-only GRPO on all five benchmarks.

Ablation results comparing the backbone, SFT, episode-only GRPO, and VTCA across five cowork benchmarks. — **Controlled training comparison.** Both RL methods improve on SFT, with additional gains from verifier-grounded turn credit.

Inside the workspaces

Files, tools, and outcomes that stay connected

Recorded examples from the manuscript show how professional work spans multiple artifacts.

Recorded agent execution

Quarterly executive handoff

An agent checks five regional revenue rows, registers a vendor, updates the Q2 access policy, retires the archived Q1 policy, and validates five exports.

Revenue workbookVendor registryAccess policyExecutive report

4 goal groups passedArchived score: 0.946. The final report retains a database lock wait and a stale export alias as follow-ups. Validation covers artifact structure and content; visual review was unavailable.

A broad skill foundation

Professional roles meet executable capabilities

The synthesis resources include 21,444 app skills, 5,918 file skills, and 36,064 job skills across 317 roles. File relations and compatible operations help assemble executable workspaces.

Treemaps of app skills, file skills, and job skills by category. — Largest categories in each skill library; smaller categories are grouped as Other.

Reference

Cite WorkRover

If this work is useful to your research, please cite the manuscript.

BibTeX

@misc{he2026workrover,
  title = {WorkRover: Scaling Long-Horizon Cowork Agent Training
           Via Recursive Environment Evolution},
  author = {Hang He and Linting Wang and Guanming Liu and Bo Mao
            and Jinxiu Liu and Yiran Suo and Maosen Zhou and Zhiheng Xi
            and Guanting Dong and Zhou Shao and Rui Zheng and Jie Zhou
            and Chengcheng Wan and Tao Gui and Qi Zhang and Xipeng Qiu
            and Xuanjing Huang},
  year = {2026},
  url = {https://workrover.github.io/}
}

Download paper ↓Website source ↗

Scaling Long-Horizon Cowork Agent Training
Via Recursive Environment Evolution