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

The Core Claim

Agent training typically relies on hand-built environments tied to specific benchmarks. This creates a scaling problem: realistic workflows involve many tasks that naturally emerge from the world itself, not tasks designed backward from a benchmark. AgentMercury flips this: synthesize rich, persistent business worlds first—complete with entities, services, tools, and cross-service rules—and let diverse tasks emerge organically. The payoff is substantial: agents trained on 4,783 auto-generated business environments improve on both enterprise workflows and out-of-domain tasks spanning reasoning, coding, and scientific computing.

How It Works

AgentMercury constructs executable environments from high-level business scenarios (e.g., "a retail company managing inventory across multiple locations"). Instead of designing individual tasks, the framework instantiates a persistent world: real entities (products, customers, warehouses), services (payment processing, shipping), tools (APIs, databases), and invariants (constraints that must hold across services, like "inventory can't go negative"). Once the world exists, diverse tasks naturally arise—fulfilling orders, handling refunds, managing stock—without explicit task engineering.

The authors built 4,783 such environments spanning 14 industries and 50 countries. These serve as training substrates for reinforcement learning. Critically, these environments were not designed to target any evaluation benchmark; they were built from business scenarios alone.

Why It Matters

Scale and realism: Prior work treats each benchmark as a separate, isolated task-environment pair. AgentMercury decouples environment construction from task definition, enabling rapid scaling to thousands of realistic worlds that reflect how businesses actually operate.

Transfer learning: Training on business-oriented environments transfers surprisingly well. A 4B-parameter model (Qwen3.5) jumped from 12.3 to 15.7 on EnterpriseOps-GYM (an enterprise workflow benchmark) and from 45.9 to 56.0 on AIME26 (a mathematical reasoning benchmark with no business connection). This suggests that learning in rich, persistent worlds teaches generalizable agent skills—adaptability, state tracking, multi-step reasoning—that apply beyond the training domain.

Learning the construction process: A secondary finding: the authors show that environment synthesis itself can be learned. Fine-tuning a larger model (35B-A3B) on traces of how environments were constructed improved success on authoring new business scenarios from 3.3% to 83.3%. This hints that agents could eventually author their own training environments, further breaking the scaling bottleneck.

The practical significance is twofold. For enterprise AI, it means agents can train on diverse, realistic workflows at scale rather than toy problems. For agent research, it challenges the benchmark-centric paradigm: general capability emerges more robustly from persistent, emergent-task worlds than from task-specific training.