AgentMercury Turns Business Scenarios into Verifiable Worlds for Agents

Introduction

Training an agent for real work requires more than a collection of simulations built around isolated benchmark questions. Enterprise workflows typically span several systems and connect customers, orders, inventory, permissions, payments, and support tickets. An action in one service may alter the state of another. Producing enough environments that are executable, diverse, and easy to verify is therefore becoming a central infrastructure challenge for agentic reinforcement learning.

AgentMercury takes a different starting point. Rather than asking what single task an agent should solve, it first builds a business world that can persist and evolve.

Key points

• From task-centric to world-centric synthesis. Given a high-level business scenario, the framework instantiates entities, services, tools, and state. Tasks are no longer the only design target; they can be derived from the world and its possible state transitions.
• Cross-service invariants are executable. Relationships between services are represented as constraints that the environment can check. This allows the simulator to evaluate not only whether a tool call returns an answer, but also whether the action preserves business logic.
• A broad generation scope. The authors report 4,783 executable environments spanning 14 industries and 50 countries. The supplied material does not detail the configuration of every environment, but the reported scope reflects an attempt to model business diversity rather than one benchmark domain.
• Reported transfer beyond the source scenarios. Without targeting the evaluation benchmarks during generation, policies trained on AgentMercury environments improved Qwen3.5-4B from 12.3 to 15.7 on EnterpriseOps-GYM, and from 45.9 to 56.0 on AIME26. The abstract also mentions gains across reasoning, coding, scientific computing, and tool use, but does not provide the full experimental breakdown.
• The builder can be trained as well. Fine-tuning Qwen3.5-35B-A3B on environment-construction traces increased its success at authoring executable worlds. The available material does not state the size of that improvement.

Why it matters

The main contribution is a different unit for organizing agent training data. Conventional synthesis often creates an input, a tool set, and an answer for one predefined task. This can produce many exercises, but the exercises may be disconnected from one another. A persistent world can support multiple goals, initial states, and action paths under the same business rules, making training closer to the open-ended nature of enterprise operations.

The approach may also make reinforcement-learning feedback more meaningful. Rewards could reflect valid state transitions, preserved cross-service constraints, and consistency after an action—not only whether the final response matches an expected answer. If this mechanism scales reliably, enterprise agents may be trained less like systems solving a queue of puzzles and more like systems operating inside a continuously changing business world.

The current material also leaves important questions open. It does not describe the full synthesis pipeline, quality-control process, baselines, compute cost, or ablation results. It is also unclear how much of the out-of-domain improvement comes from general tool-use skills. AgentMercury is therefore best viewed as a promising environment-engineering paradigm, not yet as proof that realistic business deployment has been solved.

Source: Hugging Face Daily Papers