Paper deep dive
Scaling Verifiable Environments for Long-horizon Work Agents
Jiazheng Zhang, Long Ma, Yunxian Yang, Zhiheng Xi, Zhikai Lei, Yajie Yang, Chenyang Liao, Enyu Zhou, Yang Nan, Yuchen Tian, Senjie Jin, Yibo Wang, Wei He, Boyang Liu, Jixuan Huang, Xin Guo, Zhezheng Hao, Xinbing Liang, Zhihao Zhang, Changzhi Zhou, Wiggin Zhou, Tao Gui, Qi Zhang, Xuanjing Huang, Clarenceai, Aiden Adams
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Work agents operate over digital artifacts to execute professional knowledge-intensive work, requiring training environments that support long-horizon interaction and trustworthy verification. However, hand-crafted environments incur prohibitive engineering overhead that prevents environment scaling, whereas synthesis methods sacrifice workspace complexity, realism, or grounded verifiability. To bridge this gap, we introduce WorkForge, a scalable synthesis framework for constructing verifiable work-agent environments from real-world resources. Starting from expert workflows, WorkForge first identifies the resources, decisions, and deliverables required by each workflow. It then retrieves relevant real-world files and organizes them into a workspace. WorkForge inspects the workspace to extract concrete, checkable facts about its content. These factual anchors fix which task types the workspace can support and how their outcomes can be verified. Therefore, WorkForge derives each task's instructions, solution plan, and complementary programmatic and semantic verifiers directly from these factual anchors, keeping verification traceable to observable workspace evidence. Furthermore, we construct 16.7K verifiable environments across 40 professional domains, with workspaces collectively covering 60 file types. Post-training Qwen3.5-35B-A3B-Base improves GDPVal from 45.5 to 73.6 and APEX Score from 5.0 to 21.3, while enabling Qwen3.5-27B to achieve highly competitive performance and outperform strong competitors. Our analyses confirm the efficacy of the proposed method and reveal consistent scaling behaviors across both data volume and interaction horizons.
Tags
Links
- Source: https://arxiv.org/abs/2610.04906v1
- Canonical: https://arxiv.org/abs/2610.04906v1
Trouble viewing inline? Open PDF directly →
Full Text
69,974 characters extracted from source content.
Expand or collapse full text
SCALING VERIFIABLE ENVIRONMENTS FOR LONG- HORIZON WORK AGENTS Jiazheng Zhang 1,2,† ∗ Long Ma 2,∗ Yunxian Yang 2,∗ Zhiheng Xi 2,† Zhikai Lei 2 Yajie Yang 2 Chenyang Liao 2 Enyu Zhou 2 Yang Nan 2 Yuchen Tian 1 Senjie Jin 2 Yibo Wang 1 Boyang Liu 2 Jixuan Huang 2 Xin Guo 2 Zhezheng Hao 2 Xinbing Liang 1 Zhihao Zhang 2 Changzhi Zhou 1 Wiggin Zhou 1 Tao Gui 2,† Qi Zhang 2,† Xuanjing Huang 2 Clarenceai 1,† Aiden Adams 1,† 1 LLM Department, Tencent 2 Fudan University jzzhang24, zhxi22@m.fudan.edu.cn tgui, qz@fudan.edu.cn clarenceai, aidenadams@tencent.com ABSTRACT Work agents operate over digital artifacts to execute professional knowledge- intensive work, requiring training environments that support long-horizon inter- action and trustworthy verification. However, hand-crafted environments incur prohibitive engineering overhead that prevents environment scaling, whereas syn- thesis methods sacrifice workspace complexity, realism, or grounded verifiability. To bridge this gap, we introduce WORKFORGE, a scalable synthesis framework for constructing verifiable work-agent environments from real-world resources. Starting from expert workflows, WORKFORGE first identifies the resources, de- cisions, and deliverables required by each workflow. It then retrieves relevant real-world files and organizes them into a workspace. WORKFORGE inspects the workspace to extract concrete, checkable facts about its content. These factual anchors fix which task types the workspace can support and how their outcomes can be verified. Therefore, WORKFORGE derives each task’s instructions, solu- tion plan, and complementary programmatic and semantic verifiers directly from these factual anchors, keeping verification traceable to observable workspace evi- dence. Furthermore, we construct 16.7K verifiable environments across 40 profes- sional domains, with workspaces collectively covering 60 file types. Post-training Qwen3.5-35B-A3B-Base improves GDPval from 45.5 to 73.6 and APEX Score from 5.0 to 21.3, while enabling Qwen3.5-27B to achieve highly competitive per- formance and outperform strong competitors. Our analyses confirm the efficacy of the proposed method and reveal consistent scaling behaviors across both data volume and interaction horizons. 1INTRODUCTION Autonomous agents mark a fundamental paradigm shift in large language models, transitioning foundation models from passive conversational assistants to active digital workers capable of exe- cuting complex, multi-step professional knowledge work (Drouin et al., 2024; Boisvert et al., 2024; Huang et al., 2024; Patwardhan et al., 2026; Vidgen et al., 2026). Agent assistants such as Claude Code and Codex illustrate this paradigm through iterative exploration, tool orchestration, and exe- cution over digital workspaces (Anthropic, 2025; OpenAI, 2025). Unlike conversational or API-calling routines, professional workflows center on concrete deliver- ables: users specify objectives, and the solution path is embedded within the workspace (Tang et al., 2026). Producing these deliverables requires agents to untangle multiple file dependencies and make decisions in authentic workspace artifacts (Zhou et al., 2024; Trivedi et al., 2024; Zhang et al., 2026a; Xi et al., 2026a). Cultivating agents capable of such long-horizon workflow fundamentally demands agent environments that couple authentic operational friction with trustworthy, verifiable supervision (Dou et al., 2025; Zhang et al., 2025; Sun et al., 2026; Xi et al., 2026b). ∗ Equal Contribution † Corresponding Author 1 arXiv:2610.04906v1 [cs.CL] 4 Oct 2026 vs. Real Project Tree Recommander.py Many Files Distributed Evidence Small Clean Tree Repo/ Readme.md App.py Sample.csv MdSummary Services: DB, Web, Search Method: TF-IDF + Cosine Index: Pickled Shards Agent Exploration Littleresearchisneeded ReviewReport + Findings Register Over-CuratedSparse ContextSalient Evidence Report: Service Topology, Method, Risks Register: Evidence, Impact Repo/ Artifacts/ Src/ Docs/ .github/ Readme.md Tests& CI Deploy&index Tfidf= TdfModel Index = Similarity() Cosine(q, index) Quick Start Services Method Test_fetch.py Workflow.yml DB, Arxiv, Web Index.0 Index.pkl Plan ExploreSearch CheckUpdated Plan Agent Loop Real Noise Focused Agent ScreenShots/ SyntheticWorkspace A Real-World Workspace B Agent Exploration Easy hard OneBlueprint Instruction : InspectaSoftwareProject Workflow: inspect & cross-checkResources: code, docs,configsOutput: report + findings register Workspace Complexity A: Synthetic B: Real-World Average Workspace Size 13× A B 21.5 KB 279 KB Files Per Workspace 5.4× A B 7.9 42.3 File TypeCoverage 2.4× A B 25 60 Real-World Workspace have Broader Evidence Space for Exploration and Check C C Blueprint Software Project Review Inspect a software project across its code, documentation, tests, configurations, and other project files. Cross-check evidence to understand the system, identify issues and risks, and produce a review report with a findings register. Figure 1:Comparison of synthetic and realistic workspaces. Realistic workspaces are sub- stantially more complex: they contain more files, span a broader range of file types, and distribute task-relevant evidence throughout the directory hierarchy. In contrast, synthetic workspaces are highly curated, with most task-relevant information concentrated in a small set of identifiable files. While expert-curated environments offer high fidelity, their prohibitive overhead severely impedes environment scaling (Zhou et al., 2024; Trivedi et al., 2024). Recent work addresses this bottleneck through automated environment synthesis (Song et al., 2026; Tu et al., 2026; Cheng et al., 2026; Fan et al., 2026; Zhao et al., 2026; Bai et al., 2026). However, these methods rely on synthetic environments or task-specific mock resources (Song et al., 2026; Tu et al., 2026; Bai et al., 2026), yielding toy workspaces with limited file diversity and cross-file dependencies. This reduces the exploration demands that characterize real professional work in Figure 1. Likewise, evaluation criteria not grounded in observable workspace facts risk introduc- ing hallucinated or otherwise unsupported requirements, leading to unreliable judges (Zhang et al., 2026b; Shi et al., 2026). To address this challenge, we introduce WORKFORGE, a scalable framework for constructing com- plex, verifiable work-agent environments from real-world digital resources. WORKFORGE is mo- tivated by a simple premise: professional tasks and their verification are inseparable from the workspaces in which they are performed. Workspace determines what an agent can explore, while observable workspace facts determine what can be asked and reliably verified. To operationalize this premise, WORKFORGE formalizes expert workflows as task blueprints, retrieves relevant resources, and profiles the resulting workspaces to recover file roles, dependencies, and factual anchors. These anchors jointly condition task generation and verifier design, keeping evaluation traceable to the workspace: programmatic checks verify deterministic requirements, while semantic rubrics assess open-ended deliverables against the same facts. Autonomous solvers then execute these tasks to yield long-horizon trajectories for post-training. WORKFORGE instantiates 16.7K verifiable environments across 40 professional domains, with workspaces comprising real artifacts, rich file types and cross-file dependencies. Fine-tuning Qwen3.5-35B-A3B-Base (Qwen Team, 2026) on the resulting trajectories improves performance across five benchmarks, including gains of 28.1 points on GDPval and 16.3 points on APEX. Con- trolled studies display significant benefits with increasing volume and trajectory length, and confirm the efficacy of the proposed method. Our contributions are as follows: 2 Expert Workflow Environment Synthesis Task Generation Traj. Generation Expert-curated Scene Blueprint Office Data Research Scene Blueprint DomainCategory Task intentsInteraction Workflow Resources DecisionsDeliverable Acquisition agent Search Research Preview Download Assign Done Independent pools Repository Dataset Web source Task Instruction Task package Programmatic Score Semantic Score Scored Rollout Record Interaction Trajectory & Artifacts Outcome Scores & Metadata 1 2 34 Workspace Tools Files Initial State Workspace Summary Schemas Profile Facts Instantiate Hidden Plan Programmatic Verifier Semantic Judge Workspace + Task Plan + Verifiers Task Design Verifier Module Visible Hidden Solver Agent Workspace Tools + Files Rollout Actions + Artifacts Interaction Generation Agent Exploration Mapping Env Figure 2: Overview of WORKFORGE pipeline. Conditioned on professional scene blueprints, WORKFORGE acquires and instantiates real-world resources into authentic workspaces, derives tasks and complementary verifiers, and collects scored trajectories through interaction. • We introduce WORKFORGE, a scalable framework that constructs verifiable environments from real-world resources, grounding task specifications and verification rubrics in audited workspaces. • We instantiate 16.7K agentic environments across 40 professional domains, featuring complex filesystem and long-horizon solver trajectories. • Empirical experiments demonstrate the efficacy of our method. Notably, fine-tuning on WORK- FORGE data yields substantial gains across five diverse agent benchmarks, improving GDPval by 28.1 points and APEX by 16.3 points. 2WORKFORGE To scale agent environments without compromising authenticity or verifiability, we introduce WORKFORGE, a framework for translating professional workflows into realistic, verifiable agent environments. WORKFORGE first distills domain expertise into reusable workflow specifications. It then instantiates each specification as a workspace populated with realistic artifacts and analyzes the resulting workspace evidence to ground both task construction and verification. Agents execute these tasks to produce scored trajectories for training. Figure 2 summarizes this four-stage pipeline. 2.1WORKFLOW SPECIFICATION To generate sophisticated professional tasks, we engage domain experts to distill their workflows into reusable Scene Blueprints, as these workflows identify dependencies between initial conditions, domain distinctive operations, and final deliverables. We define a collection of M blueprints as B =b j M j=1 , b j = (W j ,O j ,R j ,D j ),(1) whereW j specifies the workflow and its context;O j defines the expected deliverable;R j specifies required sources and their dependencies; andD j records domain specific decisions. For instance, a teacher can distill the workflow of testing students on a unit and providing targeted ex- planations into a scene blueprint. The blueprint specifies how to compare student responses against a standard answer and use teaching materials to group errors. It thus serves as a reusable seed for constructing workspaces and generating tasks with different assessment and teaching materials. Each blueprint serves as a seed that connects expert intent to task generation. Its resource require- ments guide source acquisition. Simultaneously, its workflow, decision points, and deliverable guide task generation using evidence from the instantiated workspace. 2.2ENVIRONMENT SYNTHESIS Environment synthesis integrates two modules, environment mapping and environment profiling. The former one maps each scene blueprint to different resource realizations. Complementing the 3 mapping process, the latter one represents each environment’s professional context, resource char- acteristics, and observable facts. For a scene blueprint b j , the stage produces Env(b j ) =(W jk ,P jk ) K j k=1 ,(2) where K j is the number of retained environments, W jk is a workspace, and P jk is its profile. Environment Mapping. Environment mapping expresses the one-to-many relationship between a scene blueprint and its concrete environments. We acquire distinct resource poolsR jk K j k=1 , each containing the files of a repository and defining a resource mapping as W jk = φ R jk (b j ),(3) where φ R jk realizes b j using the selected resource pool. An acquisition agent iteratively searches for sources, reads reference pages, downloads resources, and previews their contents to assess relevance and associate artifacts with the abstract source requirements R j . Each workspace uses one non- empty pool and retains its file organization. Environment Profiling. Environment profiling defines the description accompanying each envi- ronment as an integral part of synthesis. We represent this profile as P jk = (U jk ,F jk ,A jk ), F jk = n f (ℓ) jk o L jk ℓ=1 ,(4) where U jk summarizes the domain, file roles, and potential task directions;F jk contains workspace facts; and A jk contains each artifact schemas and statistics. In brief, the profile describes what makes each environment suitable for particular professional tasks and supplies factual anchors for task construction and subsequent solution verification. 2.3TASK GENERATION Task generation produces executable task packages with aligned task requirements and verification criteria. Task Design combines the scene blueprint and environment profile to specify the profes- sional objective, required deliverables, and a hidden solution plan. Verifier Design translates these requirements into rubrics and complementary programmatic and semantic verifiers. Task Design. Given a scene blueprint b j and the profile P jk , the task designer jointly generates the instruction q jk which agent can view and hidden solution plan h jk : (q jk ,h jk ) = (f task (b j ,P jk ),f plan (b j ,P jk )),(5) Together, the instruction and plan form an executable task grounded in the concrete contents of the real-world workspace. Outputs with invalid paths or any unsupported components are regenerated. Verification Module. We verify key factual claims in h jk against the workspace W jk and record the resulting evidence as b F jk . Along with the profile facts and task constraints, this evidence defines an atomic rubric set V jk = V prog jk ∪V sem jk . After execution, programmatic criteria check the final artifacts Y jk : S prog jk = 1 |V prog jk | X v∈V prog jk I[v(W jk ,Y jk ) = 1].(6) Semantic criteria and the task instruction q jk configure the LLM judge G sem jk , which assesses open- ended requirements using criterion-specific scoring guides and normalizes each score s v ∈ [0, 10]: S sem jk = 1 |V sem jk | X v∈V sem jk s v 10 .(7) Both verifiers are applied after execution, and their scores are retained separately. The resulting task package separates visible inputs from hidden solution plan and verifiers: T jk = (W jk ,q jk ) |z visible , P jk ,h jk , b F jk ,V jk ,G prog jk ,G sem jk |z hidden .(8) 4 Table 1: Key statistics of WORKFORGE from task, workspace and trajectory three dimensions. StatisticValue Task # Task number16,665 Avg. Rubrics21.9 Workspace Avg. Files42.3 Avg. Size279.0 KB # File Types60 Trajectory Avg. Steps31.8 Avg. Tool Calls33.8 Avg. Tokens76,485 (c) Task Score(d) Tool Call (a) Domain Tabular Data Images Tabular Data 53.31% Images 18.69% Structured Text 13.21% Code 7.27% PlainText 5.58% Audio 1.26% Structured Text Code PlainText Audio Other 0.68% (b) File Type office equity news database education gaming advertise science ecommerce business energy enterprise payment mapping work cybersecurity entertainment tool cooking medical job health hr video fitness AI ML Estate legal transportation tourism automotive manufacturing environment supply language robot insurance service quant customer architecture sale design dietengineering Figure 3: Distributions of WORKFORGE. Statistics across (a) Pro- fessional domains, (b) File types, (c) Task scores, and (d) Tool calls. 2.4TRAJECTORY GENERATION Trajectory generation turns static task package into observed agent behavior. For eachT jk , a solver agent π θ receives only the visible pair (W jk ,q jk ) and interacts with the workspace until it submits its deliverables or reaches the execution limit. We denote this process as (τ jk ,Y jk ) = Rollout(π θ ;q jk ,W jk ),(9) where τ jk records the ordered messages and tool interactions, and Y jk denotes the output artifacts. After execution, the graders G prog jk and G sem jk evaluate the final workspace and Y jk , producing the programmatic and semantic scores defined in Eq. 6 and 7. We retain the complete execution as a scored trajectory record z jk = τ jk ,Y jk ,S prog jk ,S sem jk ,μ jk , D rollout =z jk j,k ,(10) where μ jk contains the model configuration. Collectively, the four stages transform professional workflows encoded in scene blueprints into independently grounded task packages and interaction trajectories paired with verifiable execution outcomes. Table 1 reports the corpus-level scalar statis- tics, while Figure 3 summarizes its domain coverage, file types, task score, and tool call distribution. Algorithm 1 in Appendix B summarizes the integral procedures from scene blueprints to trajectories. 3EXPERIMENTS 3.1EXPERIMENTAL SETUP Baselines. We evaluate models trained on WORKFORGE-16K against three groups of baselines. Frontier Proprietary Models: Claude Opus 5.0 (Anthropic, 2026), Claude Opus 4.8, GPT-5.6 Sol (OpenAI, 2026), and Gemini 3.7 Flash (Doshi, 2026). Open-source Models (>100B): Qwen3- Coder (Qwen Team, 2025), Qwen3.5-397B-A17B (Qwen Team, 2026), DeepSeek-V3.2, DeepSeek- V4-Flash, DeepSeek-V4-Pro (DeepSeek-AI et al., 2026), Kimi-K2.5 (Kimi Team, 2026), GLM- 5.2 (Z.ai, 2026), MiniMax-M3 (MiniMax, 2026b), and MiniMax-M2.7 (MiniMax, 2026a). Open- source Models (<100B): OpenThinkerAgent-32B (Raoof et al., 2026), ClawGym-30A3 (Bai et al., 2026), EnvScaler-Qwen3-8B (Song et al., 2026), Nemotron-Terminal-32B (Pi et al., 2026), and Arctic-AWM-14B (Wang et al., 2026a). Benchmarks. We evaluate work-agent capability on five popular benchmarks.WildClaw- Bench (Ding et al., 2026) and Claw-Eval (Ye et al., 2026) emphasize long-horizon execution, multi- file interaction, and reliable tool use under repeated trials. GDPval (Patwardhan et al., 2026) and APEX-Agents (APEX) (Vidgen et al., 2026) focus on professional deliverables and workplace tool orchestration, while Agents’ Last Exam (ALE) (Sun et al., 2026) tests extended autonomous prob- lem solving with verifiable outcomes. Additional evaluation details are provided in Appendix C.4. 5 Table 2: Performance (%) of agent capability on WildClaw, ClawEval, GDPval, APEX, and Agent’s Last Exam. The best results are in bold and the second-best results are underlined within each group. Additional evaluation details are provided in Appendix C.4. Model WildClawClawEvalGDPvalAPEXALE ScorePass@1ScorePassˆ3ScoreScorePass@1ScorePass@1 Frontier Proprietary Models Claude Opus 5.064.532.482.374.291.058.542.748.326.7 Claude Opus 4.864.030.581.572.189.146.234.046.224.8 GPT-5.6 Sol 53.529.577.368.990.151.837.548.326.7 Gemini 3.7 Flash52.618.169.260.086.151.136.946.925.7 Open-source Models (>100B) Qwen3-Coder32.715.237.323.467.711.42.721.15.7 Qwen3.5-397B-A17B 38.817.160.853.079.727.412.931.111.4 DeepSeek-V3.228.215.258.349.274.217.66.326.39.5 DeepSeek-V4-Flash 59.227.674.466.483.646.330.042.322.9 DeepSeek-V4-Pro60.731.470.461.688.448.532.145.323.2 Kimi-K2.541.114.356.744.478.524.111.032.511.4 GLM-5.2 58.228.673.163.885.245.730.447.924.8 MiniMax-M356.828.670.255.686.841.226.034.617.1 MiniMax-M2.741.216.260.848.781.925.311.528.49.5 Open-source Models (<100B) OpenThinker-32B16.75.731.217.221.26.83.12.00.0 ClawGym-30A314.96.735.419.435.02.50.49.12.9 EnvScaler-8B13.73.818.78.215.04.90.80.70.0 Nemotron-Terminal-32B8.82.912.91.85.81.80.61.20.0 Agent World Model 14B13.17.626.016.424.53.60.81.10.8 Ours Qwen3.5-35B-A3B-Base18.67.635.419.945.55.00.611.43.8 + WorkForge-A3B29.920.053.341.873.621.39.228.712.4 Qwen3.5-27B 30.518.161.952.079.425.712.732.714.3 + WorkForge-27B38.523.865.455.782.429.916.333.917.1 Implementation Details. We apply supervised fine-tuning on Qwen3.5-35B-A3B-Base and Qwen3.5-27B. Both models are trained for five epochs with a global batch size of 64 and a maxi- mum sequence length of 256K tokens. We adopt a learning rate of 1× 10 −5 with cosine decay and 5% warmup. More implementation illustrations are in Appendix C. 3.2MAIN RESULTS Superior performance of WORKFORGE over other baselines. Table 2 reports the capacity of general agents across state-of-the-art competitors. We make several key observations: (1) WORK- FORGE delivers consistently competitive results across all benchmarks, establishing strong perfor- mance. Notably, WORKFORGE-A3B achieves the highest GDPval Score among prior specialized sub-100B baselines, i.e., 73.6, surpassing the previous SOTA, Clawgym-30A3, by a substantial mar- gin of 38.6 points. Similar trends are observed on WildClaw, ClawEval, APEX, and ALE. (2) Importantly, these leads do not shrink under strict pass rate: WORKFORGE-A3B reaches 12.4% Pass@1 on ALE, compared with 2.9% of ClawGym-30A3 which rarely complete these chal- lenge tasks. (3) WORKFORGE-A3B also delivers a substantial improvement over pretrained model Qwen3.5-35B-A3B-Base. These gains reflect WORKFORGE’s design principle: solving each task requires the agent to explore complex, real-world file structures and integrate evidence across mul- tiple files into its deliverables. This capability generalizes well on challenging benchmarks. WORKFORGE-27B advances capability toward the frontier.WORKFORGE-27B demonstrates that the supervision produced by our pipeline complements existing agent post-training, improving Qwen3.5-27B across all five benchmarks, with WildClaw, APEX, and GDPval Scores rising by 8.0, 4.2, and 3.0 points, respectively. Notably, on GDPval, WORKFORGE-27B reaches 82.4, surpassing Qwen3.5-397B-A17B (79.7) despite nearly 15× fewer total parameters, and narrowing the gap to frontier systems such as Claude Opus 4.8 (89.1) and GPT-5.6 Sol (90.1). 6 1K3.6K8K16K Data scale 15 20 25 30 Score (%) ALE 15.7 21.5 25.8 28.7 MiniMax M2.7 28.4 DeepSeek V3.2 26.3 1K3.6K8K16K Data scale 20 25 30 Score (%) WildClaw 21.5 24.9 27.0 29.9 DeepSeek V3.2 28.22 G1G2G3 Turn group 13 15 17 19 Score (%) ALE 13.3 14.7 17.5 G1G2G3 Turn group 22 24 26 28 Score (%) WildClaw 22.8 27.3 27.8 (a)(b) Figure 4: Scaling analysis of WORKFORGE. Performance on ALE and WildClaw with increasing (a) Training data volume and (b) Interaction length. 3.3ABLATION STUDIES % $ # $ % $ "" !! # % Figure 5: Effect of trajectory harness. We compare two training groups, each using 8k trajectories generated by Claude Code or OpenClaw, respectively. Trajectory Harness. We further compare mod- els trained on trajectories generated by two agent harnesses, Claude Code (Anthropic, 2025) and OpenClaw (Steinberger, 2026). In Figure 5, the model trained on Claude Code trajectories achieves a higher ALE score when evaluated with Claude Code (24.2 vs. 16.1). On WildClaw, this order- ing reverses: the model trained on OpenClaw tra- jectories leads by 6.2 points under the OpenClaw evaluation harness. The OpenClaw-trained model also performs better on APEX, although its advan- tage narrows to 2.7 points under the official ReAct Toolbelt. These results indicate a degree of harness- specific transfer: aligning the training and evalu- ation harnesses is associated with improved per- formance, potentially because the model acquires harness-specific tool-use conventions in addition to task knowledge. Table 3: Ablation of verification-based trajec- tory selection. Each variant uses trajectories selected by the semantic score, the program- matic score, or their combination. ALEAPEX Score Pass@1Score LLM only15.624.811.18 Programmatic14.894.810.92 Hybrid22.1111.415.39 Verification Mechanism. Recall that WORK- FORGE evaluates task deliverables with two com- ponents: programmatic criteria check exact values and observable workspace states, and LLM-based criteria assess content quality. To isolate their ef- fects on trajectory selection, we construct three 1K- trajectory training sets from the same pool of can- didate tasks. The programmatic-only and LLM- only variants select trajectories whose correspond- ing score is at least 0.7, while the hybrid variant applies the same threshold to the equally weighted mean of the two scores. All variants use the same training-set size and fine-tuning configuration. As shown in Table 3, the hybrid strategy outperforms either variant. This indicates that combining the two verification signals yields effective supervision. Workspace Authenticity. We instantiate matched workspaces from the same scene blueprints and vary only whether their resources are retrieved from the web or synthesized by an LLM. Figure 1 reveals a sharp structural gap: workspaces built from real-world resources contain 5.4× more files and 2.4× more file types, with their average footprint increasing from 21.5 KB to 279 KB. Real- world acquisition therefore enlarges both the scale and heterogeneity of the workspace, providing a richer artifact and evidence space for task construction and agent exploration. 7 Visual Layout Numerical ClaimsCross-source Reconciliation Reasoning Coverage Content Augmentation Qwen - MoE WF - A3b Qwen27b WF - 27B 30.3 25.531.4 53.5 45.6 71.059.877.578.972.0 74.164.686.482.077.4 87.172.892.286.583.4 25 35 45 55 65 75 85 95 (b) Rubric-level Pass Rate on GDPval Avg. Number of External Web Domains Visited Workspace Evidence Retrieval (%) $"!!%#%""##"#$#%#$ (a) Execution behavior on GDPval(b) Rubric-level Pass Rate on GDPval Figure 6: Execution behavior and rubric-level performance on GDPval. (a) The x-axis presents the mean number of distinct external web domains visited per task; the y-axis presents evidence retrieval within the workspace as a percentage of tool calls. Arrows connect each base model to its fine-tuned counterpart; color records reasoning length after tool errors. (b) Pass rate across five rubric categories. Details about rubric categories are described in Table 11. 3.4SCALING ANALYSIS Scaling with Data Volume. To investigate how training data volume affects work-agent perfor- mance, we train Qwen3.5-35B-A3B-Base on 1K, 3.6K, 8K, and 16K trajectories under the same recipe. As shown in Figure 4 ( a), performance improves monotonically with data volume. The model trained on 16K trajectories achieves scores of 28.7 on ALE and 29.9 on WildClaw, outper- forming its 1K-trajectory counterpart by 13.0 and 8.4 points, respectively. At this scale, the model also surpasses DeepSeek V3.2 (26.3) and MiniMax M2.7 (28.4) on ALE, as well as DeepSeek V3.2 (28.2) on WildClaw. Increasing data volume exposes the model to a broader range of workspace structures, cross-file dependencies, and execution paths. Broader coverage promotes transferable capabilities for workspace exploration and deliverable production. Scaling with Interaction Horizon. Beyond data volume, we explore how interaction horizon shapes agent behavior. We construct three subsets of 1K Claude Code trajectories, grouped by trajectory length: 0–20 turns (G1), 21–40 turns (G2), and 41–60 turns (G3). As shown in Fig- ure 4 (b), performance increases with trajectory length on both benchmarks. The model trained on G3 achieves scores of 17.5 on ALE and 27.8 on WildClaw, exceeding its G1 counterpart by 4.2 and 5.0 points, respectively. This improvement is accompanied by more sustained interaction with the environment. On WildClaw, the G3-trained model averages 24.0 productive tool calls, compared with 11.7 for the G1-trained model, and performs 38.3 distinct actions, compared with 17.0. Table 13 further shows that the G3-trained model uses a broader range of tool types and invokes workspace skills more frequently. These results suggest that longer training trajectories expose the model to more complete sequences of dependency tracing, evidence gathering, and deliverable construction, thereby promoting the sustained interaction required in complex workspaces. 3.5BEHAVIOR ANALYSIS We analyze the trajectory behavior of WORKFORGE-27B. Figure 6 (a) reveals a shift toward ex- ploration behavior: the fine-tuned model visits 6.52 external web domains per task (5.3× the base model’s average) and lifts targeted evidence retrieval from 5.9% to 16.2% of tool calls. After a failed tool call, reasoning length grows from 144 to 614 characters (over fourfold), reflecting de- liberate diagnosis and replanning rather than immediate retries. These shifts carry through to final deliverables: Figure 6 (b) shows category-score gains of 4.5–13 points across five GDPval-derived categories. Consistent with this shift toward exploration behavior, the largest gains appear in for- matting and visual presentation (+13 p) and numerical and calculation accuracy (+8.2 p), which rely on thorough investigation and careful verification. 8 4RELATED WORK Agent Environment Synthesis. Automated environment construction expands the supply of ex- ecutable settings for agent training. EnvScaler and Agent World Model synthesize tool interfaces and stateful backends (Song et al., 2026; Wang et al., 2026a), while Agent-World incorporates real- service metadata and external resources (Dong et al., 2026). Skill and intent-driven pipelines con- struct tasks and supporting workspaces for terminal and general-assistant workflows (Cheng et al., 2026; Bai et al., 2026). NexForge compiles capability requirements into task directives, then re- trieves or constructs the resources needed for execution (Zhao et al., 2026). WORKFORGE instead instantiates and profiles realistic workspaces before specifying concrete tasks. Preserving reposi- tory and dataset organization grounds professional tasks in heterogeneous materials, existing file relationships, and observable evidence. Long-Horizon Agent Data Synthesis. Agentic data synthesis expands multi-step execution through tool interactions (Prabhakar et al., 2025; Xu et al., 2025), task composition (Shi et al., 2025; Xie et al., 2025), and exploration-based task discovery (Ramrakhya et al., 2025; Chen et al., 2026; Wang et al., 2026b). Nemotron-Terminal studies generation, filtering, and long-context training (Pi et al., 2026), while OpenThoughts-Agent examines task sources and diversity to train OpenThinkerAgent-32B (Raoof et al., 2026). Executable-state alignment (Shi et al., 2026) and rubric-based artifact assessment (Starace et al., 2025; Zhuge et al., 2025) connect trajectory quality to task outcomes. WORKFORGE generates long-horizon supervision through sustained exploration, cross-file reasoning, and artifact production in content-rich workspaces. Task requirements and ver- ification criteria share the same workspace evidence, enabling complementary programmatic and semantic evaluation to select trajectories for agent post-training. 5CONCLUSION, LIMITATIONS AND FUTURE WORK Conclusion We introduce WORKFORGE, a scalable framework for constructing complex, veri- fiable work-agent environments from real-world resources. Its core principle is to use observable workspace facts as shared anchors for task construction and outcome verification. Starting from expert workflows, WORKFORGE retrieves and profiles real-world resources, then derives task spec- ifications, solution plans, and complementary programmatic and semantic verifiers from the result- ing factual anchors. The pipeline instantiated 16.7K environments across 40 professional domains, which improves the capability of Qwen3.5-35B-A3B-Base, including gains of 28.1 points on GDP- val and 16.3 points on APEX. Our analysis indicates significant gains with increasing data volume and trajectory length, while confirming the efficacy of verification mechanism. Limitations and Future WorkWORKFORGE depends on expert workflows and accessible re- alistic resources, limiting comparable grounding in proprietary or sensitive domains. Because workspace profiles, tasks, and verifiers are constructed automatically, their reliability depends on the completeness and accuracy of workspace profiling. Moreover, the current verifiers emphasize final deliverables, receive limited expert calibration, and do not directly assess intermediate agent behavior. Future work should broaden workflow and resource coverage while developing verifica- tion that is aware of processing. AI USE STATEMENT Generative AI tools were used to assist and polish manuscript writing. They were also used within WORKFORGE to generate synthetic task packages and training trajectories, as detailed in the method and experimental sections. The authors review all AI-assisted outputs and take responsibility for the final content of this work. ETHICS STATEMENT This paper introduces a scalable framework for constructing complex, verifiable work-agent envi- ronments from real-world digital resources. The methods and findings presented are intended solely for research purposes; our study does not pertain to malicious applications, unintended uses, or issues related to fairness, privacy, security, crowdsourcing, or human subject research. 9 REFERENCES Anthropic.Claude 3.7 sonnet and claude code, February 2025.URL https://w. anthropic.com/news/claude-3-7-sonnet. Anthropic. Introducing Claude Opus 5, July 2026. URL https://w.anthropic.com/ news/claude-opus-5. Fei Bai, Huatong Song, Shuang Sun, Daixuan Cheng, Yike Yang, Chuan Hao, Renyuan Li, Feng Chang, Yuan Wei, Ran Tao, Bryan Dai, Jian Yang, Wayne Xin Zhao, and Ji-Rong Wen. ClawGym: A scalable framework for building effective claw agents, 2026. URL https://arxiv.org/ abs/2604.26904. L ́ eo Boisvert, Megh Thakkar, Maxime Gasse, Massimo Caccia, Thibault Le Sellier De Chezelles, Quentin Cappart, Nicolas Chapados, Alexandre Lacoste, and Alexandre Drouin. WorkArena++: Towards compositional planning and reasoning-based common knowledge work tasks. In Ad- vances in Neural Information Processing Systems, volume 37, p. 5996–6051, 2024.doi: 10.52202/079017-0195. Aili Chen, Chi Zhang, Junteng Liu, Jiangjie Chen, Chengyu Du, Yunji Li, Ming Zhong, Qin Wang, Zhengmao Zhu, Jiayuan Song, Ke Ji, Junxian He, Pengyu Zhao, and Yanghua Xiao. DIVE: Scaling diversity in agentic task synthesis for generalizable tool use, 2026. URL https:// arxiv.org/abs/2603.11076. Zihao Cheng, Hongru Wang, Zeming Liu, Xinyi Wang, Xiangrong Zhu, Yuhang Guo, Wei Lin, Jeff Z. Pan, and Yunhong Wang. Terminal-World: Scaling terminal-agent environments via agent skills, 2026. URL https://arxiv.org/abs/2605.20876. DeepSeek-AI et al. DeepSeek-V4: Towards highly efficient million-token context intelligence, 2026. URL https://arxiv.org/abs/2606.19348. Shuangrui Ding, Xuanlang Dai, Long Xing, Shengyuan Ding, Ziyu Liu, JingYi Yang, Penghui Yang, Zhixiong Zhang, Xilin Wei, Xinyu Fang, Yubo Ma, Haodong Duan, Jing Shao, Jiaqi Wang, Dahua Lin, Kai Chen, and Yuhang Zang. WildClawBench: A benchmark for real-world, long-horizon agent evaluation, 2026. Guanting Dong, Junting Lu, Junjie Huang, Wanjun Zhong, Longxiang Liu, Shijue Huang, Zhenyu Li, Yang Zhao, Xiaoshuai Song, Xiaoxi Li, Jiajie Jin, Yutao Zhu, Hanbin Wang, Fangyu Lei, Qinyu Luo, Mingyang Chen, Zehui Chen, Jiazhan Feng, Ji-Rong Wen, and Zhicheng Dou. Agent- World: Scaling real-world environment synthesis for evolving general agent intelligence, 2026. URL https://arxiv.org/abs/2604.18292. Tulsee Doshi.Introducing Gemini 3.7 Flash, August 2026.URL https://blog. google/innovation-and-ai/models-and-research/gemini-models/ introducing-gemini-3-7-flash. Shihan Dou, Jiazheng Zhang, Jianxiang Zang, Yunbo Tao, Weikang Zhou, Haoxiang Jia, Shichun Liu, Yuming Yang, Shenxi Wu, Zhiheng Xi, Muling Wu, Rui Zheng, Changze Lv, Limao Xiong, Shaoqing Zhang, Lin Zhang, Wenyu Zhan, Rongxiang Weng, Jingang Wang, Xunliang Cai, Yueming Wu, Ming Wen, Yixin Cao, Tao Gui, Xipeng Qiu, Qi Zhang, and Xuanjing Huang. Multi-programming language sandbox for llms. In Pushkar Mishra, Smaranda Muresan, and Tao Yu (eds.), Proceedings of the 63rd Annual Meeting of the Association for Computational Linguis- tics (Volume 3: System Demonstrations), ACL 2025, Vienna, Austria, July 27 - August 1, 2025, p. 40–50. Association for Computational Linguistics, 2025. doi: 10.18653/V1/2025.ACL-DEMO.5. URL https://doi.org/10.18653/v1/2025.acl-demo.5. Alexandre Drouin, Maxime Gasse, Massimo Caccia, Issam H. Laradji, Manuel Del Verme, Tom Marty, David Vazquez, Nicolas Chapados, and Alexandre Lacoste. WorkArena: How capable are web agents at solving common knowledge work tasks? In Proceedings of the 41st Inter- national Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, p. 11642–11662. PMLR, 2024. URL https://proceedings.mlr.press/ v235/drouin24a.html. 10 Zhiyuan Fan, Tinghao Yu, Yuanjun Cai, Jiangtao Guan, Yun Yang, Dingxin Hu, Jiang Zhou, Xing Wu, Zhuo Han, Feng Zhang, and Lilin Wang. Toward scalable terminal task synthesis via skill graphs, 2026. URL https://arxiv.org/abs/2604.25727. Qian Huang, Jian Vora, Percy Liang, and Jure Leskovec. MLAgentBench: Evaluating language agents on machine learning experimentation. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, p. 20271– 20309. PMLR, 2024.URL https://proceedings.mlr.press/v235/huang24y. html. Kimi Team. Kimi K2.5: Visual agentic intelligence, 2026. URL https://arxiv.org/abs/ 2602.02276. MiniMax. MiniMax M2.7: Early echoes of self-evolution, March 2026a. URL https://w. minimax.io/news/minimax-m27-en. MiniMax. MiniMax M3: Frontier coding, 1M context, native multimodality—all in one model, June 2026b. URL https://w.minimax.io/blog/minimax-m3. OpenAI.Introducing codex, May 2025.URL https://openai.com/index/ introducing-codex/. OpenAI. GPT-5.6: Frontier intelligence that scales with your ambition, July 2026. URL https: //openai.com/index/gpt-5-6/. Tejal Patwardhan, Rachel Dias, Elizabeth Proehl, Grace Kim, Michele Wang, Olivia Watkins, Si- mon Fishman, Marwan Aljubeh, Phoebe Thacker, Laurance Fauconnet, Natalie Kim, Samuel Miserendino, Gildas Chabot, David Li, Patrick Chao, Michael Sharman, Alexandra Barr, Amelia Glaese, and Jerry Tworek.GDPval: Evaluating AI model performance on real- world economically valuable tasks. In International Conference on Learning Representations, 2026. URL https://proceedings.iclr.c/paper_files/paper/2026/file/ 290c2430f91912204f30bbcc990f1d-Paper-Conference.pdf. Renjie Pi, Grace Lam, Mohammad Shoeybi, Pooya Jannaty, Bryan Catanzaro, and Wei Ping. On data engineering for scaling LLM terminal capabilities, 2026. URL https://arxiv.org/ abs/2602.21193. Akshara Prabhakar, Zuxin Liu, Ming Zhu, Jianguo Zhang, Tulika Manoj Awalgaonkar, Shiyu Wang, Zhiwei Liu, Haolin Chen, Thai Hoang, Juan Carlos Niebles, Shelby Heinecke, Weiran Yao, Huan Wang, Silvio Savarese, and Caiming Xiong. APIGen-MT: Agentic pipeline for multi-turn data generation via simulated agent-human interplay. In Advances in Neural Information Processing Systems, volume 38, 2025. URL https://proceedings.neurips.c/paper_files/ paper/2025/file/5e3661f7fe4c8ac5652d62eb3d3c96ea-Paper-Datasets_ and_Benchmarks_Track.pdf. Qwen Team. Qwen3-Coder: Agentic coding in the world, July 2025. URL https://qwen.ai/ blog?id=qwen3-coder. Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026. URL https://qwen. ai/blog?id=qwen3.5. Ram Ramrakhya, Andrew Szot, Omar Attia, Yuhao Yang, Anh Nguyen, Bogdan Mazoure, Zhe Gan, Harsh Agrawal, and Alexander Toshev. Scaling synthetic task generation for agents via exploration, 2025. URL https://arxiv.org/abs/2509.25047. Negin Raoof, Richard Zhuang, Marianna Nezhurina, Etash Guha, Atula Tejaswi, Ryan Marten, Charlie F. Ruan, Tyler Griggs, Alexander Glenn Shaw, Hritik Bansal, E. Kelly Buchanan, Artem Gazizov, Reinhard Heckel, Chinmay Hegde, Sankalp Jajee, Daanish Khazi, Emmanouil Kouk- oumidis, Xiangyi Li, Hange Liu, Shlok Natarajan, Harsh Raj, Nicholas Roberts, Ethan Shen, Nishad Singhi, Michael Siu, Ashima Suvarna, Hanwen Xing, Patrick Yubeaton, Robert Zhang, Leon Liangyu Chen, Xiaokun Chen, Steven Dillmann, Saadia Gabriel, Xunyi Jiang, Anurag Kashyap, Boxuan Li, Yein Park, Minh Pham, Sujay Sanghavi, Lin Shi, Ke Sun, Yixin Wang, Zhiwei Xu, Erica Zhang, Siyan Zhao, Wanjia Zhao, Jenia Jitsev, Alex Dimakis, Benjamin Feuer, 11 and Ludwig Schmidt. OpenThoughts-Agent: Data recipes for agentic models, 2026. URL https://arxiv.org/abs/2606.24855. Dingfeng Shi, Jingyi Cao, Qianben Chen, Weichen Sun, Weizhen Li, Hongxuan Lu, Fangchen Dong, Tianrui Qin, King Zhu, Minghao Liu, Jian Yang, Ge Zhang, Jiaheng Liu, Changwang Zhang, Jun Wang, Yuchen Eleanor Jiang, and Wangchunshu Zhou. TaskCraft: Automated generation of agentic tasks, 2025. URL https://arxiv.org/abs/2506.10055. Kou Shi, Zun Wang, Qisheng Su, Shiting Huang, Ziao Zhang, Zhen Fang, Qingnan Ren, Jin Liu, Yu Zeng, Yiming Zhao, Lin Chen, Zehui Chen, and Feng Zhao. FACET: Preserving source intent and executable state in terminal task synthesis, 2026. URL https://arxiv.org/ abs/2608.18580. Xiaoshuai Song, Haofei Chang, Guanting Dong, Yutao Zhu, Ji-Rong Wen, and Zhicheng Dou. En- vScaler: Scaling tool-interactive environments for LLM agent via programmatic synthesis, 2026. URL https://arxiv.org/abs/2601.05808. Giulio Starace, Oliver Jaffe, Dane Sherburn, James Aung, Jun Shern Chan, Leon Maksin, Rachel Dias, Evan Mays, Benjamin Kinsella, Wyatt Thompson, Johannes Heidecke, Amelia Glaese, and Tejal Patwardhan. PaperBench: Evaluating AI’s ability to replicate AI research. arXiv preprint arXiv:2504.01848, 2025. doi: 10.48550/arXiv.2504.01848. Peter Steinberger. Introducing OpenClaw, January 2026. URL https://openclaw.ai/blog/ introducing-openclaw. Yiyou Sun, Xinyang Han, Weichen Zhang, Yuanbo Pang, Tianyu Wang, Yuhan Cao, Yixiao Huang, Chris Duroiu, Haoyun Zhang, Jeffrey Lin, Weishu Zhang, Tyler Zeng, Ying Yan, Bo Liu, Hanson Wen, Mingyang Xu, Xiaoyuan Liu, Zimeng Chen, Weiyan Shi, Amanda Dsouza, Vincent Sunn Chen, Dawn Song, Patrick Bryant, Carl Boettiger, Yamini Rangan, Bradley Rothenberg, Kyle Steinfeld, Arvind Rao, Tapio Schneider, Georgios Yannakakis, Laure Zanna, Kaan Ozbay, Ida Sim, Tarek Zohdi, George Em Karniadakis, Jack Gallant, Teresa Head-gordon, et al. Agents’ last exam. arXiv preprint arXiv:2606.05405, 2026. doi: 10.48550/arXiv.2606.05405. Zirui Tang, Xuanhe Zhou, Yumou Liu, Linchun Li, Yukai Wu, Weizheng Wang, Hongzhang Huang, Wei Zhou, Jun Zhou, Jiachen Song, Shaoli Yu, Jinqi Wang, Zihang Zhou, Hongyi Zhou, Yuting Lv, Jinyang Li, Jiashuo Liu, Ruoyu Chen, Chunwei Liu, GuoLiang Li, Jihua Kang, and Fan Wu. Workspace-Bench 1.0: Benchmarking AI agents on workspace tasks with large-scale file dependencies, 2026. URL https://arxiv.org/abs/2605.03596. Harsh Trivedi, Tushar Khot, Mareike Hartmann, Ruskin Manku, Vinty Dong, Edward Li, Shashank Gupta, Ashish Sabharwal, and Niranjan Balasubramanian. AppWorld: A controllable world of apps and people for benchmarking interactive coding agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, p. 16022–16076, 2024. doi: 10. 18653/v1/2024.acl-long.850. Dunwei Tu, Hongyan Hao, Hansi Yang, Yihao Chen, Yi-Kai Zhang, Zhikang Xia, Yu Yang, Yue- qing Sun, Xingchen Liu, Furao Shen, Qi Gu, Hui Su, and Xunliang Cai. ScaleEnv: Scaling environment synthesis from scratch for generalist interactive tool-use agent training, 2026. URL https://arxiv.org/abs/2602.06820. Bertie Vidgen, Austin Mann, Abby Fennelly, John Wright Stanly, Lucas Rothman, Marco Burstein, Julien Benchek, David Ostrofsky, Anirudh Ravichandran, Debnil Sur, Neel Venugopal, Alan- nah Hsia, Isaac Robinson, Calix Huang, Olivia Varones, Daniyal Khan, Michael Haines, Austin Bridges, Jesse Boyle, Koby Twist, Zach Richards, Chirag Mahapatra, Brendan Foody, and Osvald Nitski. APEX-Agents, 2026. Zhaoyang Wang, Canwen Xu, Boyi Liu, Yite Wang, Siwei Han, Zhewei Yao, Huaxiu Yao, and Yux- iong He. Agent world model: Infinity synthetic environments for agentic reinforcement learning, 2026a. URL https://arxiv.org/abs/2602.10090. Ziyi Wang, Yuxuan Lu, Yimeng Zhang, Pei Chen, Ziwei Dong, Jing Huang, Jiri Gesi, Xianfeng Tang, Chen Luo, Qun Liu, Yisi Sang, Hanqing Lu, Manling Li, Jin Lai, and Dakuo Wang. Tra- jectory2Task: Training robust tool-calling agents with synthesized yet verifiable data for complex 12 user intents. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics, p. 44021–44044, 2026b. doi: 10.18653/v1/2026.acl-long.2037. Zhiheng Xi, Jixuan Huang, Chenyang Liao, Baodai Huang, Jiaqi Liu, Honglin Guo, Yajie Yang, Rui Zheng, Junjie Ye, Jiazheng Zhang, Wenxiang Chen, Wei He, Yiwen Ding, Guanyu Li, Zehui Chen, Zhengyin Du, Xuesong Yao, Yufei Xu, Jiecao Chen, Tao Gui, Zuxuan Wu, Qi Zhang, Xuan- jing Huang, and Yu-Gang Jiang. AgentGym-RL: An open-source framework to train LLM agents for long-horizon decision making via multi-turn RL. In International Conference on Learning Representations, 2026a. URL https://openreview.net/forum?id=ZgCCDwcGwn. Zhiheng Xi, Dingwen Yang, Jiaqi Liu, Jixuan Huang, Honglin Guo, Baodai Huang, Tinggang Chen, Qi Zhang, Zhonghang Lu, Chenyu Liu, Jiajun Sun, Jiazheng Zhang, Dingwei Zhu, Xin Guo, Junzhe Wang, Zhihao Zhang, Yuming Yang, Junjie Ye, Minghe Gao, Dongrui Liu, Jiaming Ji, Guohao Li, Tao Gui, Qi Zhang, and Xuanjing Huang. AgentGym2: Benchmarking large language model agents in de-idealized real-world environments. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 44451–44479. Association for Computational Linguistics, 2026b. doi: 10.18653/v1/2026.acl-long.2058. URL https://aclanthology.org/2026.acl-long.2058/. Jingxu Xie, Dylan Xu, Xuandong Zhao, and Dawn Song. AgentSynth: Scalable task generation for generalist computer-use agents, 2025. URL https://arxiv.org/abs/2506.14205. Zhangchen Xu, Adriana Meza Soria, Shawn Tan, Anurag Roy, Ashish Sunil Agrawal, Radha Poovendran, and Rameswar Panda. TOUCAN: Synthesizing 1.5m tool-agentic data from real- world MCP environments, 2025. URL https://arxiv.org/abs/2510.01179. Bowen Ye, Rang Li, Qibin Yang, Yuanxin Liu, Linli Yao, Hanglong Lv, Zhihui Xie, Chenxin An, Lei Li, Lingpeng Kong, Qi Liu, Zhifang Sui, and Tong Yang. Claw-Eval: Towards trustworthy evaluation of autonomous agents, 2026. Z.ai. GLM-5.2: Built for long-horizon tasks, June 2026. URL https://z.ai/blog/glm-5. 2. Jiazheng Zhang, Wenqing Jing, Zizhuo Zhang, Zhiheng Xi, Shihan Dou, Rongxiang Weng, Jiahuan Li, Jingang Wang, Mingxu Chai, Shibo Hong, Tao Gui, and Qi Zhang. Two minds better than one: Collaborative reward modeling for LLM alignment. CoRR, abs/2505.10597, 2025. doi: 10. 48550/ARXIV.2505.10597. URL https://doi.org/10.48550/arXiv.2505.10597. Jiazheng Zhang, Ziche Fu, Junrui Shen, Yunbin Zhao, Yunke Zhang, Zhiheng Xi, Long Ma, Chenxin An, Zhihao Zhang, Shichun Liu, Dingwei Zhu, Shihan Dou, Shaofan Liu, Han Li, Wiggin Zhou, Aiden Adams, Tao Gui, Fei Huang, Qi Zhang, and Xuanjing Huang.En- tropy polarity in reinforcement fine-tuning: Direction, asymmetry, and control, 2026a. URL https://arxiv.org/abs/2605.11775. Jiazheng Zhang, Ziche Fu, Zhiheng Xi, Wenqing Jing, Mingxu Chai, Wei He, Guoqiang Zhang, Chenghao Fan, Chenxin An, Wenxiang Chen, Zhicheng Liu, Haojie Pan, Dingwei Zhu, Tao Gui, Qi Zhang, and Xuanjing Huang. AgentV-RL: Scaling reward modeling with agentic verifier. In Findings of the Association for Computational Linguistics: ACL 2026, p. 23078–23100. Asso- ciation for Computational Linguistics, 2026b. doi: 10.18653/v1/2026.findings-acl.1156. URL https://aclanthology.org/2026.findings-acl.1156/. Jiarong Zhao, Zhikai Lei, Zhiheng Xi, Rui Zheng, Hang Yan, Jie Zhou, Qin Chen, and Liang He. NexForge: Scaling agent capabilities through requirement-driven task synthesis for LLMs, 2026. URL https://arxiv.org/abs/2607.14186. Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. WebArena: A realistic web environment for building autonomous agents. In International Conference on Learning Rep- resentations, 2024. URL https://proceedings.iclr.c/paper_files/paper/ 2024/file/4410c0711e9154a7a2d26f9b3816d1ef-Paper-Conference.pdf. 13 Mingchen Zhuge, Changsheng Zhao, Dylan R. Ashley, Wenyi Wang, Dmitrii Khizbullin, Yun- yang Xiong, Zechun Liu, Ernie Chang, Raghuraman Krishnamoorthi, Yuandong Tian, Yangyang Shi, Vikas Chandra, and J ̈ urgen Schmidhuber. Agent-as-a-judge: Evaluate agents with agents. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, p. 80569–80611. PMLR, 2025. URL https: //proceedings.mlr.press/v267/zhuge25a.html. APPENDIX ADOMAIN TAXONOMY AND DISTRIBUTION A.1DOMAIN TAXONOMY WorkForge organizes 16,665 SFT tasks into 40 domains, as shown in Figure 7. The taxonomy spans common workplace activities such as office productivity, data management, business operations, finance, and communication, while also covering software, scientific, media, healthcare, security, and infrastructure-oriented tasks. The included workflows range from document and data processing to technical investigation and domain-specific decision making. O ffi ce Productivity 8.36% Data & Database 8.21% Business & Enterprise 7.94% Finance & Payments 7.42% Advertising & Marketing 5.90% AI & Machine Learning 4.21% Science & Research 3.86% Education & Learning 3.84% Communication & Messaging 3.80% Movies & Video 3.34% Social Media 2.89% Entertainment 2.83% Tools & Utilities 2.68% Jobs & HR 2.66% Monitoring & Observability 2.46% Search & Information Retrieval 2.46% Other 2.38% Gaming 2.35% Cybersecurity & Cryptography 2.08% Media & News 1.73% Commerce & eCommerce 1.65% Devices & IoT 1.63% Medical 1.46% Visual Recognition & Image Processing 1.42% Storage & File Management 1.27% Events 1.27% Health & Fitness 1.16% Music & Audio 0.98% Food & Cooking 0.97% Translation & Language 0.96% Location & Mapping 0.94% Email 0.74% Energy 0.68% Sports 0.65% Travel & Tourism 0.60% Text Analysis & NLP 0.54% Weather 0.51% Transportation 0.48% Logistics & Supply Chain 0.35% SMS & Noti fi cations 0.34% WorkForge 40 domains 16,665 tasks (a) Task domain taxonomy 0200400600800100012001400 Number of tasks Movies & Video Communication & Messaging Education & Learning Science & Research AI & Machine Learning Advertising & Marketing Finance & Payments Business & Enterprise Data & Database Office Productivity 557 (3.34%) 633 (3.80%) 640 (3.84%) 644 (3.86%) 701 (4.21%) 983 (5.90%) 1,237 (7.42%) 1,324 (7.94%) 1,368 (8.21%) 1,394 (8.36%) (b) Top 10 domains by task count Figure 7: Domain taxonomy and task distribution of the WorkForge SFT corpus. The left panel shows all 40 domains, and the right panel reports the ten most represented domains. A.2DOMAIN DISTRIBUTION Table 4 lists the complete distribution. The four most represented domains account for 31.9% of the corpus, and the top ten account for 56.9%. The remaining 30 domains contribute 43.1% of all tasks. The distribution combines repeated coverage of broadly applicable workplace domains with coverage across specialized applications. BPIPELINE PSEUDOCODE AND PROMPT TEMPLATES B.1PIPELINE PSEUDOCODE Algorithm 1 presents the end-to-end construction procedure and follows the notation introduced in Section 2. 14 Table 4: Complete domain distribution. Counts are computed over 16,665 SFT tasks. # DomainCount%# DomainCount% 1 Office Productivity1,394 8.36 21 Commerce & eCommerce275 1.65 2 Data & Database1,368 8.21 22 Devices & IoT271 1.63 3 Business & Enterprise1,324 7.94 23 Medical243 1.46 4 Finance & Payments1,237 7.42 24 Visual Recognition & Image Processing237 1.42 5 Advertising & Marketing983 5.90 25 Storage & File Management212 1.27 6 AI & Machine Learning701 4.21 26 Events211 1.27 7 Science & Research644 3.86 27 Health & Fitness194 1.16 8 Education & Learning640 3.84 28 Music & Audio163 0.98 9 Communication & Messaging633 3.80 29 Food & Cooking162 0.97 10 Movies & Video557 3.34 30 Translation & Language160 0.96 11 Social Media482 2.89 31 Location & Mapping156 0.94 12 Entertainment471 2.83 32 Email123 0.74 13 Tools & Utilities447 2.68 33 Energy113 0.68 14 Jobs & HR443 2.66 34 Sports108 0.65 15 Monitoring & Observability410 2.46 35 Travel & Tourism100 0.60 16 Search & Information Retrieval410 2.46 36 Text Analysis & NLP90 0.54 17 Other396 2.38 37 Weather85 0.51 18 Gaming391 2.35 38 Transportation80 0.48 19 Cybersecurity & Cryptography346 2.08 39 Logistics & Supply Chain59 0.35 20 Media & News289 1.73 40 SMS & Notifications57 0.34 Algorithm 1: Authority-Separated Task Construction and Trajectory Generation in WORKFORGE Input: Expert-authored professional workflow specifications; resource-acquisition interfaceA ground ; solver policy π θ Output: Grounded task packagesT ; scored rollout corpusD rollout 1 T ← ∅; D rollout ← ∅; // Phase 1: Workflow Specification 2 Obtain expert-authored Scene BlueprintsB =b j M j=1 ; 3 foreach Scene Blueprint b j ∈B do // Phase 2: Environment Synthesis 4 R jk K j k=1 ← GroundResourcePools(b j ;A ground ); 5foreach resource pool R jk do 6W jk ← Instantiate(R jk ); 7P jk ← Explore(W jk ;b j ); // Phase 3: Task Generation and Verification Design 8Synthesize task instruction q jk and solution plan h jk from b j and P jk ; 9 b F jk ← VerifyFacts(h jk ;W jk ,P jk ); 10(V prog jk ,V sem jk )← BuildRubrics(q jk ,h jk ,P jk , b F jk ); 11 V jk ←V prog jk ∪V sem jk ; 12Construct G prog jk fromV prog jk and G sem jk from q jk andV sem jk ; 13 T jk ← ((W jk ,q jk ), (P jk ,h jk , b F jk ,V jk ,G prog jk ,G sem jk )); 14 T ←T ∪T jk ; // Phase 4: Trajectory Generation 15(τ jk ,Y jk )← Rollout(π θ ;q jk ,W jk ); 16S prog jk = 1 |V prog jk | X v∈V prog jk I[v(W jk ,Y jk ) = 1], S sem jk = 1 |V sem jk | X v∈V sem jk s v 10 ; 17 D rollout ←D rollout ∪(τ jk ,Y jk ,S prog jk ,S sem jk ,μ jk ); 18 returnT ,D rollout ; CEXPERIMENTAL, TRAINING, AND EVALUATION DETAILS C.1SUPERVISED FINE-TUNING SETUP We perform full-parameter supervised fine-tuning on 16,665 trajectories from WORKFORGE-16k. The training trajectories are generated by GLM 5.3 with the Claude Code and OpenClaw harnesses. 15 We train Qwen3.5-35B-A3B-Base and the post-trained Qwen3.5-27B model for five epochs using the same optimization recipe. Table 5 reports the training configuration. Table 5: Supervised fine-tuning configuration used for the main models and training-data ablations. ConfigurationValue OptimizerAdamW Adam coefficientsβ 1 = 0.9, β 2 = 0.95 Peak learning rate1× 10 −5 Weight decay0.1 Gradient clipping1.0 Learning-rate scheduleCosine decay to 1× 10 −7 Warmup ratio5% of total steps Training epochs5 Global batch size64 sequences Micro-batch size1 sequence per GPU Maximum sequence length262,144 tokens Numerical precisionBF16 mixed precision Sequence packingEnabled Dynamic batchingEnabled Gradient checkpointingEnabled C.2TRAINING INFRASTRUCTURE AND DATA FORMATTING Training uses the verl supervised fine-tuning trainer with Ray 2.46, PyTorch 2.7.1, CUDA 12.8, Transformers 5.10.4, and FlashAttention 2.8.1. We use FSDP with Ulysses sequence parallelism of size four. The Qwen3.5-35B-A3B-Base runs use 64 NVIDIA H20 96GB GPUs, and the Qwen3.5- 27B runs use 128 GPUs. Optimizer states are offloaded to CPU, while model parameters remain on GPU. Training examples are rendered incrementally from cumulative conversation prefixes so that the reasoning content associated with each assistant turn is preserved. The loss is applied according to the per-message assistant loss mask stored with each trajectory. Variable-length examples are packed without padding, with sequence boundaries retained through cumulative sequence-length indices. Dynamic batching limits each GPU batch by its token count, with a maximum budget of 262,144 tokens. C.3INFERENCE CONFIGURATION We serve local models with vLLM and use the same decoding configuration for the WORKFORGE models and the re-evaluated open-source baselines in Table 2. Table 6 lists the shared settings. The maximum context length is 262,144 tokens, whereas the 32,768-token limit applies to each gen- erated model response. For models whose native context window is shorter than 262,144 tokens, we use YaRN position extrapolation to extend the context window to the shared limit. Specifi- cally, EnvScaler-Qwen3-8B, Arctic-AWM-14B, Nemotron-Terminal-32B, and OpenThinkerAgent- 32B use the Qwen3 architecture with the YaRN settings in Table 7. Table 6: Decoding configuration used for model evaluation. ParameterValue Temperature0.7 Top-p0.9 Top-k20 Maximum generated tokens per response32,768 Maximum context length262,144 16 Table 7: Shared YaRN configuration for the Qwen3-based baselines. ParameterValue Maximum position embeddings262,144 RoPE base (θ)1,000,000 RoPE scaling typeYaRN Scaling factor6.4 Original maximum position embeddings40,960 C.4BENCHMARK PROTOCOLS All results in Table 2 are obtained through our evaluation runs on WildClawBench, Claw-Eval, GDPval, APEX-Agents, and Agents’ Last Exam. Table 8 summarizes their metrics and sampling protocols. Pass@1 denotes the fraction of tasks passed in a single rollout under each benchmark’s success criterion. Claw-Eval uses three independent rollouts per task and reports Passˆ3, the fraction of tasks passed in all three rollouts. Unlike Pass@3, which requires at least one successful rollout, Passˆ3 measures consistency across repeated runs. All other benchmarks use one rollout per task. WildClawBench and Claw-Eval use OpenClaw; GDPval and ALE use Claude Code; APEX-Agents uses its official ReAct Toolbelt agent. Table 9 lists the harnesses, agent-level wall-clock limits, and interaction budgets. Table 8: Benchmark-specific metrics and sampling protocols. Benchmark Evaluation setReported metricsRollouts/task WildClawFull setScore, Pass@11 Claw-EvalFull setScore, Passˆ33 GDPval201 text-only tasksRubric score1 APEXFull setScore, Pass@11 ALEFull setScore, Pass@11 Table 9: Evaluation harnesses and per-task agent execution budgets. Benchmark HarnessWall-clock limit Interaction limit WildClawOpenClaw24,000 sNot set Claw-EvalOpenClaw3,600 s50 turns GDPvalClaude Code14,400 s300 turns APEXReAct Toolbelt14,400 s250 steps ALEClaude Code28,800 sUnlimited C.5GDPVAL EVALUATION SUBSET We evaluate all models on the same 201-task subset of GDPval’s official 220-task evaluation set, excluding 19 multimodal tasks. Tasks are aligned using the evaluation payload’s caseid, which corresponds to the official taskid. We report the official rubric score rather than the win-or-tie rate. Table 10 lists the excluded tasks and their occupations. C.6GDPVAL RUBRIC-LEVEL ANALYSIS We group GDPval criteria into five deliverable-quality categories and compute the pass rate within each category. For category c and model m, let I c denote its criteria and z (m) j ∈ 0, 1 indicate whether criterion j is satisfied. The category pass rate is R (m) c = P j∈I c z (m) j |I c | .(11) 17 Table 10: Multimodal tasks excluded from the GDPval evaluation set. Row numbers refer to the official 220-task set. Row Task IDOccupation 12 38889c3b-e3d4-49c8-816a-3c8e5313aba Audio and Video Technicians 13 f85e58-bc9f-4a2-806d-87edeabb1b81 Audio and Video Technicians 14 4b894ae3-1f23-4560-b13d-07ed1132074e Audio and Video Technicians 56 e222075d-5d62-4757-ae3c-e34b0846583b Film and Video Editors 57 c94452e4-39cd-4846-b73a-ab75933d1ad7 Film and Video Editors 58 75401f7c-396d-406d-b08e-938874ad1045 Film and Video Editors 59 a941b6d8-4289-4500-b45a-f8e4fc94a724 Film and Video Editors 79 4d1a8410-e9c5-4be5-ab43-c55563c594c First-Line Supervisors of Office & Admin Support 102 b9665ca1-4da4-4f9-86f2-40b9a8683048 Industrial Engineers 111 46fc494e-a24f-45ce-b099-851d5c181fd4 Mechanical Engineers 127 3baa0009-5a60-4ae8-ae99-4955cb328f3 News Analysts, Reporters, and Journalists 134 e6429658-4de1-42d-a9e0-2d2b9b02fb10 Nurse Practitioners 147 57b2cdf2-ad62-4591-a91-aad489740320 Private Detectives and Investigators 149 a46d5cd2-55fe-48fa-a4c6-6aaf6b9991b5 Private Detectives and Investigators 153 e4f664ea-0e5c-4e4e-a0d3-a87a33da947a Producers and Directors 171 11593a50-734d-4449-b5b4-f8986a133fd8 Real Estate Sales Agents 172 94925f49-36bc-42da-b45b-61078d329300 Real Estate Sales Agents 188 f2986c1f-2bbf-4b83-bc93-624a9d617f45 Pharmacists 203 fe0d3941-e32c-4bf1-a643-b566d2b4cb3c Sales Representatives, Wholesale & Manufacturing We report 100R (m) c as a percentage and compute each gain as 100(R (SFT) c − R (Base) c ) percentage points. Table 11 reports the number and share of criteria assigned to each category together with the resulting pass rates. Table 11: GDPval rubric pass rates across five deliverable-quality categories. CategoryCriteria ShareBaseSFT∆ Visual / layout quality3714.2% 74.1% 87.1%+13 p Numerical claims5235.9% 64.6% 72.8% +8.2 p Cross-source reconciliation1031.2% 86.4% 92.2% +5.8 p Reasoning / argumentation2002.2% 82.0% 86.5% +4.5 p Content coverage7,718 86.6% 77.4% 83.4% +6.0 p DADDITIONAL ABLATION ANALYSES The training-data ablations use Qwen3.5-35B-A3B-Base with the supervised fine-tuning recipe in Table 5. D.1DATA VOLUME SCALING We train models at four data scales: 1K, 3.6K, 8K, and 16K trajectories. For each target scale, we independently sample the requested number of trajectories from the full corpus; subsets at different scales are not nested. The 16K setting uses all 16,665 trajectories. Figure 4(a) reports the results on ALE and WildClaw. D.2TRAJECTORY-LENGTH SCALING We construct the turn-scaling subsets from trajectories generated with the Claude Code harness. Trajectories are grouped into three disjoint intervals by their number of turns, and we randomly 18 sample 1,000 trajectories from each interval (Table 12). Figure 4(b) reports performance on ALE and WildClaw, while Table 13 summarizes tool-use behavior on WildClaw. Table 12: Training subsets for trajectory-length scaling. GroupTurns per trajectoryTrajectories G10–201,000 G221–401,000 G341–601,000 Table 13: Tool-use behavior under trajectory-length scaling on WildClaw. Longer-turn supervision increases exploration intensity, breadth, and productive execution. Group Productive calls Skill discovery Tool calls Distinct actions Tool types used G111.721.0%17.717.03.3 G223.121.9%38.333.03.6 G324.033.3%42.038.34.0 D.3VERIFICATION-FILTER ABLATION We construct three 1,000-trajectory training sets from the same pool of completed rollouts using the filtering rules in Table 14. The LLM-only variant selects trajectories by the semantic score S sem , the programmatic-only variant uses S prog , and the hybrid variant uses their equally weighted mean. Each variant applies a threshold of 0.7 to its selection score. The downstream results on ALE and APEX are reported in Table 3. Table 14: Trajectory selection rules for the verification-filter ablation. VariantSelection ruleTrajectories LLM onlyS sem ≥ 0.71,000 Programmatic only S prog ≥ 0.71,000 Hybrid0.5S sem + 0.5S prog ≥ 0.71,000 EDATA EXAMPLES E.1CROSS-REGION FACTOR-PREMIUM ANALYSIS This Finance & Payments task asks a quantitative strategist to assess factor premia and prepare an investment-committee briefing. WorkForge constructs the workspace from the public Kaggle dataset nikitamanaenkov/famafrench-factors-and-portfolios, a local mirror of the Kenneth French data library with an August 2025 vintage. The example connects workspace ex- ploration, evidence-grounded verification, cross-region analysis, and the production of three linked deliverables. E.1.1TASK INSTRUCTION Task Instruction You are a quantitative strategist at a multi-asset manager. The investment committee is weighing a new factor-tilt allocation and wants evidence on whether the classic equity factors still earn premia -- and where (US vs developed vs emerging markets). Using the firm’s local mirror of the Kenneth French data library, build the analysis and a decision-ready briefing. Assess each major factor’s long-run and recent performance, compare premia across 19 regions over a fair common window, and recommend which factors and regions deserve capital (and which to avoid). Be explicit about data caveats that affect your conclusions. Deliver: - results/factor_briefing.pptx -- committee briefing with findings and a clear recommendation - results/factor_analysis.xlsx -- supporting performance tables - results/factor_returns_tidy.csv -- the cleaned return series behind your analysis Explore the workspace to determine structure, coverage, and pitfalls yourself. E.1.2WORKSPACE AND SOURCE MATERIALS The workspace contains 297 CSV files in 16 directories, with a maximum nesting depth of six. It includes factor series, sorted portfolios, research breakpoints, and industry returns, organized into US, Developed, Emerging, Industry, and Other branches. Momentum data also appear outside the US factor directory. Workspace Layout workspace/ |-- data/ | |-- US/ | | |-- Factors/ [5 files] | | |-- Portfolios/ | | | |-- Bivariate/ [49 files] | | | ‘-- Univariate/ [18 files] | | ‘-- Research_Breakpoints_Data/ [8 files] | |-- Developed/ | | |-- Factors/ [36 files] | | ‘-- Portfolios/ [114 files] | |-- Emerging/ | | |-- Factors/ [2 files] | | ‘-- Portfolios/ [7 files] | |-- Industry/ [24 files] | ‘-- Other/ [34 files] ‘-- results/ [initially empty] Individual files contain descriptive preambles and multiple frequency blocks. For example, the US three-factor file stacks monthly observations and annual returns in the same CSV: Source Excerpt: F-FResearchDataFactors.csv This file was created using the 202508 CRSP database. The 1-month TBill rate data until 202405 are from Ibbotson Associates. Starting from 202406, the 1-month TBill rate is from ICE BofA US 1-Month Treasury Bill Index. ,Mkt-RF,SMB,HML,RF 192607, 2.89, -2.55, -2.39, 0.22 192608, 2.64, -1.14, 3.81, 0.25 ... 202508, 1.85, 3.87, 4.41, 0.38 Annual Factors: January-December 1927, 29.47, -2.46, -3.75, 3.12 1931, -44.84, 3.07, -13.69, 1.07 1933, 56.49, 44.61, 31.28, 0.30 Other files declare missing-value sentinels of−99.99 or−999. The US files identify CRSP as their source, whereas Developed and Emerging files identify Bloomberg. The solver must separate fre- quency blocks, handle missing values, reconcile factor names such as Mom and WML, and determine overlapping valid observations across regions. 20 E.1.3WORKSPACE EVIDENCE AND VERIFICATION CRITERIA Workspace inspection produces 12 usable fact records with computation references. One record cor- rects the assumption that international series share a July 1990 start: the Developed series begins in July 1990, but the Emerging five-factor file begins in July 1989. The final factual criterion references this corrected record. Selected Workspace Fact Records vf01: US three-factor monthly series starts at 192607. First row: 192607, 2.89, -2.55, -2.39, 0.22. vf03: The US three-factor CSV contains a monthly block followed by an annual block. vf07: File headers declare missing values as -99.99 or -999. vf09: US files use CRSP; Developed and Emerging files use Bloomberg. vf12 (corrected): Developed starts at 199007; Emerging five-factor data start at 198907. Computation: inspect the first data rows of the two regional factor files. The task has 23 criteria: eight programmatic checks and 15 semantic criteria. Programmatic checks assess deliverable validity and specific reported facts. Semantic criteria assess analysis, data han- dling, evidence-to-recommendation consistency, and deliverable quality. Selected criteria are shown below. Selected Verification Criteria DC-2 | Programmatic | file_valid The analysis workbook is valid and contains at least one populated sheet. FA-4 | Programmatic | value_in_context | fact_ref: vf12 The Emerging market 5-factor series start (198907) is correctly distinguished from the Developed start of 199007. Expected value: 198907; context keywords: Emerging, 1989, start. AD-2 | Semantic Compare regional premia over a common window, rather than each region’s full history. HC-1 | Semantic Separate monthly and annual blocks so annual returns do not enter monthly statistics. HC-2 | Semantic Treat -99.99 and -999 as missing values rather than numerical returns. HC-3 | Semantic Discuss the CRSP/Bloomberg source difference when interpreting regional comparisons. DQ-3 | Semantic Produce a tidy long-format CSV with date, region, factor, and return fields. E.1.4TRAJECTORY EXCERPTS GLM 5.3 executes the task with OpenClaw in 1,317 seconds, producing 121 messages and 60 tool calls. Early exploration checks both ends of the source files and then compares regional schemas, provenance, and date coverage. Rollout Excerpts: Inspection and Data Handling [9] Assistant Now let me inspect the actual file formats -- French CSVs are notoriously messy (preamble text, multiple stacked tables, annual sections, footers, -99.99 missing codes). [15] Assistant Key pitfalls already visible: - US is CRSP (1926+), Developed/Emerging are Bloomberg (1989/1990+) - Emerging has ‘-99.99‘ missing RMW/CMA early - whitespace-padded values - momentum is a separate file (‘Mom‘ for US, ‘WML‘ for intl) - stacked monthly+annual tables Let me check the toolchain and the Emerging RMW/CMA validity start. 21 The solver checks factor-level availability next. Emerging RMW becomes available in July 1991 and CMA in July 1992, making July 1992 the binding start for the cross-region five-factor panel. Rollout Excerpt: Common-Window Construction [32] Assistant Clean panel built. The binding constraint for a fair all-factor, all-region window is Emerging CMA starting 1992-07, so the common window is 1992-07 to 2025-08 (398 months). Now the analytics engine. Tool call: write /project/scripts/analytics.py Compute factor performance analytics from results/factor_returns_tidy.csv. Outputs (results/tables/ * .csv) consumed by the xlsx/pptx builders, plus results/_headline.json with key numbers for the briefing. 22