ZICQ
中 Log in / Sign up
Newsroom Agentic #AgentMercury #Agent Training #AI Framework #Business Scenarios #Reinforcement Learning

AgentMercury Framework Released: Scalable Environment Synthesis for Business-Centric AI Agents

Avatar of Mr.Xu

By Mr.Xu

Published:

中文阅读 (Chinese) English Version

Summary:AgentMercury is a novel framework for synthesizing executable environments from high-level business scenarios, aiming to provide AI agents with more realistic training data. By instantiating a persistent world with entities, services, tools, and cross-service invariants, AgentMercury enables the natural emergence of diverse tasks and interaction trajectories, addressing the limitations of traditional task-centric training paradigms. Experiments demonstrate that policies trained on AgentMercury-g


Key Breakthroughs

The release of the AgentMercury framework marks a significant shift in the paradigm of AI agent training. Unlike traditional methods that rely on manually constructed environments or task-specific synthesis, AgentMercury achieves breakthroughs through the following:

  • High-Level Business Scenario-Driven Environment Generation: By generating environments from high-level business scenarios, AgentMercury provides AI agents with training data that closely resembles the real world.
  • Emergent Task Diversity: The framework supports the natural emergence of diverse tasks, rather than being limited to predefined tasks and benchmarks, thus enabling more comprehensive training of agent adaptability.
  • Scalability: AgentMercury constructs 4,783 executable environments across 14 industries and 50 countries, demonstrating its strong scalability and generalization capabilities.

Technical Highlights

  1. Persistent World Construction: By constructing a persistent world with entities, services, tools, and cross-service invariants, AgentMercury provides AI agents with a dynamic and complex environment.
  2. Task Diversity: The framework supports the natural emergence of diverse tasks, allowing agents to be trained across a wider range of tasks.
  3. Performance Improvement: The Qwen3.5-4B model trained in AgentMercury environments shows a 27.6% improvement on EnterpriseOps-GYM and a 22.0% improvement on AIME26 benchmarks.
  4. Learnability: The construction process of AgentMercury itself can be learned, with the Qwen3.5-35B-A3B model achieving a success rate increase from 3.3% to 83.3% after fine-tuning.

Industry Impact

The release of the AgentMercury framework provides new ideas and methods for AI agent training, particularly in enterprise workflows and cross-domain applications. Its scalability and generalization capabilities enable AI agents to better adapt to the complex and ever-changing real-world environment, thereby enhancing their performance in practical applications. Additionally, AgentMercury offers a powerful tool for AI researchers and developers, driving further advancements in AI agent training technology.

Developer Recommendations

  • Explore New Application Scenarios: Developers can leverage the AgentMercury framework to explore its application potential in different industries and fields.
  • Optimize Training Processes: Use the environments generated by AgentMercury to optimize the training processes of AI agents and improve model performance.
  • Focus on Learnability: Further research on the learnability of AgentMercury’s construction process can enhance agent performance in complex tasks.

Source: Hugging Face Daily Papers (2026-08-21)

— END —

Tags: #AgentMercury #Agent Training #AI Framework #Business Scenarios #Reinforcement Learning

Community Comments

Loading live comments and annotations…