AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale

Minbyul Jeong; Chanwoong Yoon

cs.CLAug 21, 2026

AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale

Authors: Minbyul Jeong, Chanwoong Yoon

Organizations: Meridian Intelligence Global Inc. · University of Massachusetts Amherst

Abstract

Agents learn to act through interaction with environments, yet the environments used for training are often manually constructed or synthesized around predefined tasks and benchmarks. This task-centric paradigm makes it difficult to scale environments that reflect realistic and evolving workflows where diverse tasks can naturally emerge from the underlying world. We introduce AgentMercury, a scalable framework for synthesizing executable environments from high-level business scenarios. Rather than constructing an environment for a specific task, AgentMercury first instantiates a persistent world with entities, services, tools, state, and executable cross-service invariants, from which diverse tasks and interaction trajectories can subsequently emerge. We construct 4,783 executable environments spanning 14 industries and 50 countries, and use them as training substrates for reinforcement learning. Despite being generated without targeting the evaluation benchmarks, policies trained on these business-oriented environments improve substantially on both enterprise workflows and out-of-domain benchmarks spanning reasoning, coding, scientific computing, and tool use. In our experiments, Qwen3.5-4B improves from 12.3 to 15.7 on EnterpriseOps-GYM and from 45.9 to 56.0 on AIME26 after training on AgentMercury environments. We further show that the construction process itself can be learned: fine-tuning Qwen3.5-35B-A3B on construction traces increases executable-world authoring success from 3.3% to 83.3% on held-out business scenarios. These results show that scenario-grounded environments can provide useful and generalizable learning signals beyond benchmark-specific training, while their construction can itself become a learnable capability.

Figures & tables

Figure 1: AgentMercury synthesizes diverse, scalable, and verifiable business environments at scale. (a) The synthesized environment spans 14 industries and 50 countries, covering 4,783 environments with anonymized to prevent hallucination. (b) Comparison of environment scale and multi-service depth with existing agent environments ( Trivedi et al., 2024 ; Wu et al., 2026 ; Shi et al., 2026 ; Dong et al., 2026 ) ; AgentMercury occupies the regime that is both large-scale and multi-service. (c) Each environment contains rich, stateful structure, including multiple services, tools, state tables, and deterministic SQL-based verifiers that enforce cross-service constraints.

Figure 2: From scenario-grounded world construction to agent interaction. (A) Planet transforms a high-level business scenario $\sigma$ into an executable world $w$ , constructing its company identity, service graph, state schema, seeded initial state $s_{0}$ , and world-level invariants $\mathcal{R}$ . A task is then instantiated from the resulting world through task-specific state seeding and a rubric. (B) Within the constructed world, a policy acts from the interaction history, producing actions that deterministically transition the environment from $s_{t}$ to $s_{t+1}$ , while observations $\omega_{t+1}$ are exposed through the observation interface. The resulting trajectory is evaluated post-episode using the task rubric and, where applicable, the world-level invariants. The formulation thus separates world construction, task instantiation, interaction, and evaluation.

Model	Teams	CSM	Email	ITSM	Calendar	HR	Drive	Hybrid	Avg.
GPT-5 $\dagger$	26.3	36.4	49.0	18.9	41.3	17.9	34.0	23.5	30.9
Gemini-2.5-Pro $\dagger$	39.3	11.6	31.1	13.9	12.5	4.9	27.0	19.6	19.9
Kimi-K2-Thinking $\dagger$	30.0	7.1	51.0	12.2	15.4	8.2	39.6	15.7	22.4
Qwen3-4B (Think) $\dagger$	24.0	3.8	38.4	5.6	5.8	7.1	21.9	15.8	15.3
Qwen3-30B (Think) $\dagger$	22.0	5.4	51.9	6.7	18.3	7.6	25.7	15.7	19.1
Qwen3-235B (Think) $\dagger$	28.0	4.7	38.1	9.3	15.7	7.8	23.8	17.7	18.1

Table 1: Result on EnterpriseOps-Gym ( Malay et al., 2026 ) . We evaluate whether training on environments and tasks synthesized by AgentMercury improves agent performance across model scales and policy optimization methods. Results are reported for both Qwen3.5-4B and Qwen3.5-35B-A3B models, with GRPO ( Shao et al., 2024 ) and SAO ( Hou et al., 2026 ) considered as alternative RL algorithms.

\dagger

signifies that the result are derived from EnterpriseOps-Gym .

Model	AIME26	HMMT	LCB	SciCode	Tau3- Airline	Tau3- Retail	Tau3- Telecom	BFCL	GPQA- Diamond
Qwen3.5-4B	45.9 $\pm$ 1.7	28.5 $\pm$ 1.2	36.6 $\pm$ 2.8	22.6 $\pm$ 0.1	48.8 $\pm$ 2.1	70.4 $\pm$ 1.0	92.5 $\pm$ 1.0	30.3 $\pm$ 1.8	76.5 $\pm$ 0.7
Qwen3.5-4B + GRPO + Ours	56.0 $\pm$ 1.8	35.4 $\pm$ 1.6	44.0 $\pm$ 1.2	25.7 $\pm$ 0.7	58.7 $\pm$ 0.8	73.6 $\pm$ 1.3	91.9 $\pm$ 1.3	31.7 $\pm$ 1.5	77.5 $\pm$ 0.6
Qwen3.5-35B-A3B	91.0 $\pm$ 0.3	77.0 $\pm$ 0.9	74.3 $\pm$ 2.2	29.7 $\pm$ 1.5	39.1 $\pm$ 10.6	52.7 $\pm$ 22.1	49.1 $\pm$ 49.8	31.1 $\pm$ 1.3	82.8 $\pm$ 3.0
Qwen3.5-35B-A3B + GRPO + Ours	91.9 $\pm$ 2.1	80.0 $\pm$ 0.7	79.0 $\pm$ 0.5	29.9 $\pm$ 1.7	52.5 $\pm$ 5.2	57.1 $\pm$ 8.3	68.9 $\pm$ 20.5	42.5 $\pm$ 0.3	85.0 $\pm$ 0.5
Qwen3.5-35B-A3B + SAO + Ours	92.2 $\pm$ 1.9	83.3 $\pm$ 3.1	78.6 $\pm$ 1.0	28.3 $\pm$ 1.9	50.9 $\pm$ 8.1	55.7 $\pm$ 10.2	65.5 $\pm$ 23.8	42.1 $\pm$ 0.1	84.0 $\pm$ 1.0

Table 2: Out-of-domain evaluation across mathematical reasoning, coding, scientific computing, tool use, and agentic benchmarks. We report mean

\pm

standard deviation over three independent evaluation runs (

N=3

). Qwen3.5-4B and Qwen3.5-35B-A3B are evaluated before and after RL training on tasks synthesized by AgentMercury .

Figure 3: Out-of-domain benchmark performance across training checkpoints for Qwen3.5-4B + GRPO + Ours . The dashed line denotes the base-model performance, while error bars show the standard deviation over three independent evaluation runs. The gradual improvement across heterogeneous benchmarks indicates that the policy learned from AgentMercury environments acquires transferable capabilities beyond the training environments.

Figure 4: Training dynamics of Qwen3.5-4B + GRPO + Ours . We report the raw reward, response length, truncated-response ratio, and degenerate-response ratio over training steps. The increasing reward together with decreasing truncation and near-zero degeneration indicates that policy improvement is not accompanied by substantial response collapse or pathological generation.

Model	Zero-shot	Recipe
Claude Opus 4.8	83.3%	80.0%
GPT-5.4	66.7%	66.7%
DeepSeek-V4-Pro	83.3%	86.7%
GLM-5.2	80.0%	76.7%
Mean (API models)	78.3%	77.5%
Qwen3.5-35B-A3B	3.3%	20.0%

Table 3: Authoring executable worlds from high-level business briefs. We report oracle-pass rates over 30 held-out briefs. A world is counted as successful only when it passes all 12 structural validators. Zero-shot provides only the business brief, while Recipe additionally provides an invariant digest and a trimmed construction exemplar.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Overview of the synthesized environments, sampled training footprint, and task lifecycle in AgentMercury . (A) The full corpus contains 43,300 investigation tasks instantiated from 4,326 synthetic company environments, spanning 2,287 industries, 842 tools, 222 state tables, and 148 service combinations. (B) Across 200 RL training steps, the agent observes 3,200 tasks (7.4% of the full task corpus) while retaining broad coverage of environments, industries, tools, state tables, and service combinations through shuffled sampling. (C) Each training instance follows a structured lifecycle from a high-level business scenario to executable world construction, task instantiation, multi-step tool interaction, deterministic post-episode verification, and reward-based policy optimization. The resulting framework separates world construction from task sampling, allowing a single synthesized environment to support multiple seeded tasks while providing verifiable learning signals for agent training.

Statistic	Value
Tasks	43,300
Executable worlds / companies	4,326
Task seeds per world	10
Unique industries	2,287
Unique tools	842
Unique state tables	222

Appendix

Table 4: Composition of the synthesized task corpus used for RL. Statistics are computed over all 43,300 tasks.

Statistic	Full	Trained	Coverage
Tasks	43,300	3,200	7.4%
Environments	4,326	2,313	53.5%
Industries	2,287	1,438	62.9%
Tools	842	638	75.8%
State tables	222	137	61.7%
Service combinations	148	93	62.8%

Appendix

Table 5: Coverage of the tasks that received gradients during the reported RL run relative to the full synthesized corpus. These statistics used only for GRPO algorithm.

Figure 6: Structural analysis of the synthesized task space. (a) A t-SNE projection of task descriptions based on TF-IDF features shows substantial textual clustering, but the clusters do not correspond to meaningful task metadata such as domain, persona, or task shape. (b) In contrast, projecting the same tasks using their exposed tool sets reveals clear domain-specific structure, indicating that the primary source of task variation lies in the executable tool surface rather than in task language. (c) Distribution of nearest-neighbor cosine similarity in the TF-IDF space, showing that although the overall task space is broad, most tasks still have a highly similar counterpart. (d) The number of exposed tools takes only five values, $\{10,14,18,22,26\}$ , revealing a discrete task-shape lattice induced by the core tool set and the number of additional resources.

Parameter	Value
Tensor parallelism	2
Pipeline / context / expert parallelism	1 / 1 / 1
Sequence parallelism	Enabled
Recomputation	Full, uniform, 1 layer
Attention backend	FlashAttention
Attention softmax precision	FP32

Appendix

Table 6: Model-parallelism and memory configuration used for Qwen3.5-4B policy optimization.

Parameter	Value
Max response length	16,384
SGLang context length	24,576
Rollout temperature	1.0
End-of-turn token	<\|im_end\|>
Partial rollout	Enabled
Reasoning / tool parser	Qwen3 / Qwen3-Coder

Appendix

Table 7: Rollout and fully-asynchronous execution configuration.

Parameter	Value
Maximum MCP turns	20
Maximum generation tokens / turn	8,192
Arena backend	Local (in-process)
Probe tool	Calendar
Environment creation throttle	8
Ready timeout	600 s

Appendix

Table 8: Configuration of the executable MCP environments used during policy optimization.

Parameter	Value
Reward implementation	AgentMercury rl_reward.py
Reward aggregator	Fraction
Behavior penalty	Enabled
Reward judge	DeepSeek-V4-Flash
Judge temperature	0.0
Judge max tokens	4,000

Appendix

Table 9: Reward and verification configuration used during RL training.

Parameter	Value
Advantage estimator	GRPO
Surrogate	Decoupled-PPO ( icepop )
$\epsilon$ -clip / $\epsilon$ -clip-high	0.20 / 0.28
Entropy coefficient	0
GRPO std normalization	Disabled
Per-token loss	Enabled

Appendix

Table 10: GRPO and optimizer configuration for Qwen3.5-4B.

Parameter	Value
Base model	Qwen3.5-4B
Training corpus	AgentMercury investigation corpus
Number of tasks	43,300
Seeds per environment	10

Appendix

Table 11: Model and corpus configuration for the Qwen3.5-4B RL experiment.

Figure 7: Training dynamics of Qwen3.5-35B-A3B with GRPO and AgentMercury . We report the raw reward, response length, truncated-response ratio, and degenerate-response ratio throughout training. The increasing reward is accompanied by a steady reduction in truncation, while degenerate responses remain negligible.

Figure 8: Training dynamics of Qwen3.5-35B-A3B with SAO and AgentMercury . We report the raw reward, response length, truncated-response ratio, and degenerate-response ratio throughout training. SAO exhibits a sustained reward increase together with a substantial reduction in truncation, while degenerate responses remain close to zero.

LLM Agent Training

Large language model agent training teaches models to plan, use tools, and learn from multi-step interaction with environments.

See all 586 papers

EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL
May 18, 2026Minrui Xu, Zilin Wang, Mengyi DENG +12Tool-Augmented Language Model Agents Agentic RL
LARK, HKUST (GZ) · University of Cambridge · UCL +2
Agent-World: Scaling Real-World Environment Synthesis for Evolving General Agent Intelligence
Apr 20, 2026Guanting Dong, Junting Lu, Junjie Huang +17Lifelong Learning Agents AI Agent Benchmarks
Renmin University of China · ByteDance Seed
EnvCraft: Synthesizing Executable Environments in Agentic RL for Claw-like Agent
Sep 4, 2026Yirong Zeng, Shen You, Jinhang Feng +8LLM Agent Training Agentic RL
Harbin Institute of Technology, SCIR Lab · Huawei Technologies Co., Ltd · Peking University +1

May 18, 2026J/K move · Enter open · S save

EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL

Minrui Xu, Zilin Wang, Mengyi DENG, Zhiwei Li+11

LARK, HKUST (GZ) · University of Cambridge · UCL+2

Tool-Augmented Language Model Agents Agentic RL LLM Agent Trajectory Synthesis LLM Agents LLM Agent Training RL for Tool Use

Open paper AskPDF arXiv

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

τ^2

-Bench and VitaBench. By fully automating both environment construction and trajectory synthesis, EnvFactory provides a scalable, extensible, and robust foundation for Agentic RL.

Abstract

LLM Agent Training

Explore similar work

EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL

Agent-World: Scaling Real-World Environment Synthesis for Evolving General Agent Intelligence

EnvCraft: Synthesizing Executable Environments in Agentic RL for Claw-like Agent