The paradigm of LLMs has rapidly shifted from passive language interfaces to autonomous Claw-like agents that execute long-horizon tasks across stateful workspaces. While Agentic Reinforcement Learning (Agentic RL) provides a promising path to optimize these agents, its scaling is heavily bottlenecked by the severe scarcity of interactive training environments. Existing synthetic environments are strictly limited to tool-calling endpoints, rendering them insufficient for accommodating the end-to-end real-world demands of claw-like agents. To bridge this gap, we introduce EnvCraft, an automated framework for synthesizing executable environments and scalable training data. Specifically, EnvCraft employs an environment synthesis engine to build sandbox-isolated workspaces, alongside a topology-aware data generation engine to produce coherent task trajectories. Overall, we synthesize 139 interactive environments comprising approximately 20K complex tasks for Agentic RL training. Experiments on Qwen3/3.5 models (8B-32B) show that our method yields gains of up to +11.9% on Claw-style benchmarks and +8.0% on general tool-use benchmarks, with concurrent reductions in inference token cost. The results confirm that synthesized executable environments provide robust and generalizable learning signals for training.
Figures & tables
Figure 1: An illustration of the agent interaction loop in RL. Environment fundamentally comprises environment documentation, explicit state definitions, and a suite of tool interfaces.
Environment
# Envs
# Tasks
Claw Support
Env Domain
Stateful
Reward
AutoForge ( Cai et al. 2025 )
10
1,078
✗
Consumer Web
DB Schema
LLM Judge
EnvScaler ( Song et al. 2026 )
191
9,000
✗
General
Class Attr.
Rule Check
ScaleEnv ( Tu et al. 2026 )
16
2,560
✗
General
DB Schema
Rule Check
AWM ( Wang et al. 2026 )
1,000
10,000
✗
General
DB Schema
Code & LLM
EnvFactory ( Xu et al. 2026 )
85
2,575
✗
General
DB Schema
Code Check
EnvCraft (Ours)
139
19,777
✓
General
Workspace
Code Check
Table 1: Comparison of synthetic executable environments for Agentic RL. Unlike existing works limited to generalist tool endpoints, EnvCraft provides stateful, workspace-level infrastructure for training Claw-like agents with the largest task scale.
Figure 2: Distribution of synthesized environments across coarse-grained scenario categories.
Figure 3: Environment and task synthesis pipeline in EnvCraft. Given natural-language descriptions, the Environment Synthesis Engine generates executable sandboxes (state + tools + docs); the Data Synthesis Engine samples tool chains and synthesizes RL training tasks (intent sequence + initial state + verifier).
Env/Data Types
Passed
Raw
Pass Rate
Environment
General Tool Env
70
203
34.48%
Claw Env
69
121
57.02%
# Total
139
324
42.90%
Data Entry
General Tool Data
14,215
21,975
64.69%
Table 2: Statistics of Environment and Data from EnvCraft, along with pass rate in quality auditing.
Model
PinchBench
ClawEval
Score
Token
Score
Token
Advanced Models
Claude Opus 4.6
69.90
12.30
80.60
20.40
Kimi-2.5
54.60
14.00
66.60
17.80
GPT 5.4
75.70
11.50
78.30
15.30
MiniMax-m2.7
65.40
11.60
71.80
17.80
Table 3: Main results. Score denotes the average task score (%), and Token denotes the average token usage per trajectory (K). ±⋅ denotes the standard deviation over three runs. Bold values indicate the best score within each backbone group.
Model
BFCL-v3 Multi-Turn
τ2 -bench
Base
Miss- Func
Miss- Param
Long- Context
Overall
Retail
Airline
Telecom
Overall
Advanced Models
Gemini-3-Pro-Pre.
64.50
60.00
54.50
64.00
60.75
75.90
80.50
91.00
82.47
Claude Sonnet 4.5
69.00
65.00
52.50
59.00
61.37
72.40
72.00
84.90
76.43
Qwen3-235B
60.00
35.00
34.00
54.00
45.75
71.90
58.60
47.30
59.27
Kimi-K2-Instruct
57.50
35.00
42.00
49.00
45.88
70.60
56.50
65.80
64.30
Table 4: Performance comparison of advanced models on multi-turn tool-use benchmarks (BFCL-v3 Multi-Turn and τ2 -bench). All scores are reported as percentages. “-” indicates unreported results in the original paper.
Figure 4: Training dynamics on EnvCraft-Claw environments. Test PinchBench scores are reported throughout RL, showing consistent and stable improvements for both (a) Qwen3-8B and (b) Qwen3.5-9B as training progresses.
Figure 5: Training dynamics on the BFCL-v3 Base subset. We plot the Base score of Qwen3-8B, Qwen3-32B, and Qwen3.5-9B against RL training steps on EnvCraft.
Figure 6: Comparison of PinchBench performance and average token consumption per task before and after EnvCraft-Claw training. Higher task scores and lower token costs indicate better inference efficiency.
Appendix figures & tables21 assets
Supplementary material from the paper’s appendix.
Appendix
No.
Fine-Grained Scenario
# Envs
Scenario Description
1
Messaging & Communication
5
APIs for instant messaging, mobile messaging, email, SMS, and other communication services, supporting session management, message sending and receiving, and user-status tracking.
2
Document & Note Management
3
APIs for document-management and personal note-taking systems, supporting document creation and editing, version control, note organization, and information retrieval.
3
Development Tools
1
Jupyter Notebook server APIs supporting the creation, execution, editing, and lifecycle management of computational notebooks.
4
Scheduling & Reminders
2
APIs for scheduling and reminder systems, supporting calendar-event management, timed reminders, recurrence configuration, and notification delivery.
5
Monitoring & Logging
3
APIs for network monitoring, model monitoring, and system-log management, supporting device-status detection, log collection, alerting, and performance-metric tracking.
6
Server & Application Management
1
Node.js application-server management APIs supporting application deployment, start and stop operations, process management, and runtime status monitoring.
Appendix
Table 5: Fine-grained general tool-use scenarios and their corresponding environment descriptions.
No.
Fine-Grained Scenario
# Envs
Scenario Description
10
Ground Transport
2
Car-rental and ride-hailing platform APIs supporting vehicle leasing, ride requests, driver and vehicle matching, route planning, and fare calculation.
Hotel-booking APIs supporting hotel and room search, availability management, price calculation, booking creation, modification, and cancellation.
13
Vehicle Management
3
APIs for vehicle maintenance, motor-vehicle registration, and vehicle control, covering maintenance scheduling, registration records, and in-vehicle system operations.
14
Banking & Trading
3
APIs for retail banking, stock trading, and quantitative-trading bots, supporting account management, transfers, market-data retrieval, order placement, and automated trading.
15
Corporate Finance
1
Corporate financial-reporting APIs supporting financial-statement generation, multi-source data aggregation, accounting-period management, and compliance reporting.
Appendix
Table 6: Fine-grained general tool-use scenarios and their corresponding environment descriptions (continued).
No.
Fine-Grained Scenario
# Envs
Scenario Description
18
Online Shopping & Cart
2
APIs for online-shopping and retail platforms, supporting product browsing, cart management, promotional discounts, checkout, and order creation.
APIs for inventory and order management, supporting inventory updates, order-status transitions, fulfillment tracking, and stock reconciliation.
21
Marketing & Reviews
2
APIs for product promotion and review systems, supporting campaign configuration, promotional-rule management, user reviews, ratings, and feedback moderation.
22
Task & Issue Tracking
3
APIs for task tracking, to-do management, and software-testing workflows, supporting assignment, prioritization, status transitions, and defect tracking.
23
Workflow & Job Scheduling
2
APIs for workflow-management and job-scheduling systems, supporting process definition, dependency configuration, job execution, retry handling, and status monitoring.
Appendix
Table 7: Fine-grained general tool-use scenarios and their corresponding environment descriptions (continued).
No.
Fine-Grained Scenario
# Envs
Scenario Description
26
Social Media
1
Social-media publishing APIs supporting post creation, comment interaction, reposting, content retrieval, and follower-relationship management.
27
HR & Employee Management
1
Human-resources information-system APIs supporting employee-profile management, organizational structures, employment records, and payroll information maintenance.
28
Account & Membership
2
APIs for membership management and user registration, supporting account creation, authentication, membership-tier management, and access-control configuration.
29
Benefits & Wellness
1
Fitness-class booking APIs supporting class scheduling, member registration, coach assignment, capacity management, and booking cancellation.
30
Communication & Services
2
APIs for telecommunications and food-delivery services, supporting mobile-plan management, data-usage control, restaurant ordering, and delivery-status tracking.
31
Search & Information
1
Web-search APIs supporting keyword queries, result retrieval, result organization, search-history management, and bookmark collection.
Appendix
Table 8: Fine-grained general tool-use scenarios and their corresponding environment descriptions (continued).
No.
Fine-Grained Scenario
# Envs
Scenario Description
34
Academic Administration
3
APIs for student-information and university course-selection systems, supporting student-profile management, course registration, enrollment constraints, and grade-record maintenance.
35
Knowledge Resources
1
Library-management APIs supporting catalog search, book borrowing, reservations, renewals, return processing, and overdue-fine calculation.
36
CRM & Customer Management
1
Customer-relationship management APIs supporting customer profiles, sales-opportunity tracking, interaction histories, and communication record maintenance.
37
Reservation & Booking
1
Restaurant-reservation APIs supporting restaurant and table information management, availability queries, booking creation, and dining-schedule modification.
38
Support Ticketing
1
Customer-support ticketing APIs supporting ticket creation, assignment, status updates, priority management, escalation, and service-queue management.
Table 9: Fine-grained general tool-use scenarios and their corresponding environment descriptions (continued).
No.
Fine-Grained Scenario
# Envs
Scenario Description
1
Collaborative Office & Email Automation
3
Collaborative-office and email-automation environments supporting workflow orchestration for email-content extraction, automated social-media publishing, and event-driven email rules.
2
E-commerce & Logistics Management
5
E-commerce logistics environments supporting shipment tracking, return and exchange processing, inventory reconciliation, multi-stage logistics orchestration, and state-machine-based order management.
3
Spreadsheet Data Cleaning & Statistical Analysis
1
Spreadsheet-analysis environments supporting data cleaning, formula calculation, chart generation, pivot-table construction, and statistical analysis across workspace files.
4
Email Client Intelligent Management
1
Intelligent email-client environments supporting email sending and receiving, message organization, reply suggestion, and automatic extraction of actionable to-do items.
5
Sensitive Information Encryption & Management
1
Secure information-storage environments supporting encrypted password storage, credential retrieval, automatic form filling, and strong password generation.
6
Cross-Modal Content Transcription & Summarization
2
Cross-modal content-processing environments supporting audio- and video-to-text transcription, discussion consolidation, and multi-turn conversational summarization.
Appendix
Table 10: Fine-grained Claw-specific scenarios and their corresponding environment descriptions.
No.
Fine-Grained Scenario
# Envs
Scenario Description
10
Itinerary Disruption Dynamic Optimization
4
Dynamic itinerary-optimization environments supporting delay monitoring, automatic rebooking, linked hotel and airport-transfer adjustments, and traveler notification.
Offline travel-planning environments supporting destination organization, itinerary construction, local information management, and multi-objective route optimization.
Identity-verification and audit-defense environments supporting auditor-identity validation, anomalous-behavior detection, forgery-risk assessment, and early warning.
Experiment-analysis environments supporting experimental-configuration recording, metric aggregation, group comparison, statistical difference analysis, and result archiving.
31
Open-Source Reproduction Record Archiving
1
Reproduction-management environments supporting procedural record keeping, dependency and environment configuration, result validation, and reproducibility archiving.
32
Music & Audio Playback
1
In-vehicle audio environments supporting music search, playback control, queue manipulation, playlist management, and media-status retrieval.
Appendix
Table 13: Fine-grained Claw-specific scenarios and their corresponding environment descriptions (continued).
No.
Fine-Grained Scenario
# Envs
Scenario Description
35
Infrastructure Automated Operations
1
Infrastructure-operations environments supporting server-fault detection, diagnostic inspection, automatic recovery, service verification, and operations-record maintenance.
Table 14: Fine-grained Claw-specific scenarios and their corresponding environment descriptions (continued).
#
Constraint
Applicable Components
1
All public functions return dictionaries and map failures to structured error responses rather than propagating exceptions.
Environment, tools, CLI, storage, evaluator
2
No mutable module-level global state is permitted; all mutable data must reside in the session or environment state.
All modules
3
The data/ directory is immutable during rollout and can be read only through the repository layer.
Repository and domain modules
4
Session updates use an atomic temporary write followed by os.replace .
Storage layer
5
File locking uses platform-supported locking with a bounded timeout and an agent-safe error response.
Storage layer
6
Timestamps are computed from base_time and action_index ; wall-clock timestamps are forbidden.
Environment and domain modules
Appendix
Table 15: Code-level constraints enforced during environment generation.
Category
Example
Required Behavior
Missing binding
Session environment variable is absent
Return an error indicating that the trainer must initialize the rollout.
Session not found
The requested session does not exist
Return an agent-safe message stating that the session may have expired or been reset.
Lock timeout
Another worker holds the session lock
Request a retry without exposing the lock path.
Permission failure
The current account cannot modify an entity
Identify the affected entity and state that write permission is absent.
Invalid transition
An entity is already in the requested terminal state
Report the current state and explain why the transition is invalid.
Missing field
A required command argument is absent
Name the missing field and describe the corresponding requirement.
Appendix
Table 16: Structured error categories used in generated environments.
Hyperparameter
Value
Optimization
Optimizer
AdamW
Learning rate
1×10−6
Learning-rate scheduler
Constant
Warmup ratio
0.0
Weight decay
0.01
Appendix
Table 17: Common hyperparameters used for EnvCraft reinforcement learning. Model-specific differences are described in the main text.
Figure 7: Training dynamics of Qwen3-8B, Qwen3-32B, and Qwen3.5-9B on EnvCraft-Tool. The left panels show training performance, while the right panels report performance on 50 held-out tasks excluded from training.
Figure 8: Effect of training-data composition on Qwen3-8B. We compare general tool-use training, Claw-specific training, and sequential general-to-Claw training using checkpoint-level performance on PinchBench.
Figure 9: Retention of general tool-use capability during Claw-specific reinforcement learning. Scores are measured on the BFCL-v3 Multi-Turn Base subset at different training checkpoints.
Dimension
Score 3
Score 2
Score 1
Score 0
Task Solvability & Logic
The task is logically valid, non-trivial, and fully solvable under the initial state and environment rules. Completing it requires meaningful interaction, with no state conflicts.
The task is theoretically solvable but is either trivial, requires no meaningful interaction, or contains minor parameter mismatches that a robust agent may bypass.
The task is unsolvable because its requested actions or target state contradict the initial state, environment constants, or transition rules.
The task substantially hallucinates entities, tools, APIs, or states that are absent from the environment.
Validation-Protocol Integrity
The validation protocol is executable and completely checks the intended success conditions, without evident false-positive or false-negative paths.
The protocol is executable, but its assertions are incomplete or weak; for example, it verifies that an action occurred without checking the required final state.
The protocol contains syntax or runtime errors, references undefined variables or APIs, or is expected to fail during execution.
The validation protocol is missing or evaluates conditions unrelated to the user task.
Data Consistency & Syntactic Cleanliness
Entity names, identifiers, value types, and schemas are fully consistent across the environment state, user task, and validation protocol.
The components are semantically consistent but contain minor type or formatting mismatches that do not fundamentally alter the task objective.
Critical fields are missing, or explicit contradictions exist between the user-facing description and the underlying environment configuration.
The task record contains malformed JSON, severely corrupted text, or otherwise unusable structured data.
Appendix
Table 18: Complete rubric used by the LLM-based data-quality inspector. Each synthesized task is assigned an integer score from 0 to 3 along the three dimensions.
Figure 10: Average LLM-rubric scores of retained and filtered task candidates across task solvability, validation-protocol integrity, and data consistency. Each dimension is scored on a 0–3 scale. A candidate is retained only when it receives scores of 3, 3, and at least 2 on the three dimensions, respectively.
Figure 11: Ablation study of Claw-specific environment components on PinchBench. The three configurations progressively introduce built-in system tools and parallel skill co-synthesis. All models use the same Qwen3-8B backbone and training configuration.
Figure 12: A representative EnvCraft-Claw trajectory involving workspace inspection, cross-file record matching, rule-based computation, persistent state mutation, and output verification. For readability, repetitive intermediate reasoning and lengthy tool outputs are omitted while preserving the original action sequence and execution results.
Equipping LLMs with tool-use capabilities via Agentic Reinforcement Learning (Agentic RL) is bottlenecked by two challenges: the lack of scalable, robust execution environments and the scarcity of realistic training data that captures implicit human reasoning. Existing approaches depend on costly real-world APIs, hallucination-prone LLM simulators, or synthetic environments that are often single-turn or depend on pre-collected documents. Moreover, synthetic trajectories are frequently over-specified, resembling instruction sequences rather than natural human intents, reducing their effectiveness for RL training. We introduce EnvFactory, a fully automated framework that addresses both challenges. EnvFactory autonomously explores and verifies stateful, executable tool environments from authentic resources, and synthesizes natural multi-turn trajectories through topology-aware sampling and calibrated refinement, producing grounded queries with implicit intents. Using only 85 verified environments across 7 domains, EnvFactory generates 2,575 SFT and RL trajectories. Despite using significantly fewer environments than prior work, which are often 5 times more, EnvFactory achieves superior training efficiency and downstream performance, improving Qwen3-series models by up to +15% on BFCLv3, +8.6% on MCP-Atlas, and +6% on conversational benchmarks including τ2-Bench and VitaBench. By fully automating both environment construction and trajectory synthesis, EnvFactory provides a scalable, extensible, and robust foundation for Agentic RL.
Minrui Xu, Zilin Wang, Mengyi DENG +12
LARK, HKUST (GZ) · University of Cambridge · UCL +2
General agentic intelligence hinges on the ability to interact with diverse real-world tools to complete complex tasks, a capability fundamentally tied to the quality of interaction data. To bypass the prohibitive costs of human annotation, prevailing paradigms depend entirely on Large Language Models (LLMs) to scale the synthesis of agentic environments and tasks. However, such unconstrained generation often degenerates into biased random sampling of LLMs' internal priors, failing to capture the diversity and difficulty of real-world domains or construct high-fidelity, long-horizon tasks. In this work, we introduce Grounded Agentic Interaction Synthesis (GAIS), a framework that automates the scalable construction of diverse environments and complex tasks via a two-phase grounding mechanism. Specifically, we construct protocol-anchored environments derived from real-world Model Context Protocol (MCP) servers to ensure functional diversity and difficulty. Subsequently, we employ structure-guided planning to navigate these environments, actively enforcing logical dependencies and adversarial policies to generate complex tasks. Experiments on BFCL, τ2-Bench, and ACEBench demonstrate that GAIS-synthesized data significantly outperforms state-of-the-art baselines, enabling base models to match or even surpass their official instruction-tuned counterparts. Furthermore, GAIS exhibits superior data efficiency and scalability, achieving exceptional capabilities with significantly less data while maintaining continuous growth where baselines stagnate. Our code and dataset are publicly available at https://github.com/Eric8932/GAIS.
Wenhang Shi, Jinhao Dong, Yiren Chen +4
Renmin University of China · Peking University · Tencent
Claw-style environments support multi-step workflows over local files, tools, and persistent workspace states. However, scalable development around these environments remains constrained by the absence of a systematic framework, especially one for synthesizing verifiable training data and integrating it with agent training and diagnostic evaluation. To address this challenge, we present ClawGym, a scalable framework that supports the full lifecycle of Claw-style personal agent development. Concretely, we construct ClawGym-SynData, a diverse dataset of 13.5K filtered tasks synthesized from persona-driven intents and skill-grounded operations, paired with realistic mock workspaces and hybrid verification mechanisms. We then train a family of capable Claw-style models, termed ClawGym-Agents, through supervised fine-tuning on black-box rollout trajectories, and further explore reinforcement learning via a lightweight pipeline that parallelizes rollouts across per-task sandboxes. To support reliable evaluation, we further construct ClawGym-Bench, a benchmark of 200 instances calibrated through automated filtering and human-LLM review. Relevant resources have been released at https://github.com/ClawGym.
Fei Bai, Huatong Song, Shuang Sun +11
Gaoling School of Artificial Intelligence, Renmin University of China · IQuest Research · Beihang University