EnvFactory: 85 Simulated Environments Train a General Tool-Use Agent via Automated Trajectory Synthesis

Forum topic · 小凯 · 2026-05-19

Summary

Agentic reinforcement learning for tool use is bottlenecked by two problems: scalable executable environments and realistic training data. EnvFactory, developed by Xu, Wang, and colleagues from Huawei and MBZUAI, addresses both automatically. The system autonomously explores and validates stateful, executable tool environments from real resources, then synthesizes natural multi-turn trajectories via topology-aware sampling and calibrated refinement. Crucially, it generates queries with implicit intent rather than mechanical "step 1: click here" instruction sequences. Using only 85 validated environments across 7 domains, EnvFactory produced 2,575 training trajectories. Models trained with this data improved by up to +15% on BFCLv3, +8.6% on MCP-Atlas, and +6% on conversational benchmarks. The post's author raises open questions: whether 85 environments generalize across all domains, whether topology sampling convergence risks RL policy overfitting to a few typical paths, and how the automated framework avoids invoking real tools with side effects when interacting with live APIs.

Agentic RL training for tool-use capabilities faces two bottlenecks: scalable execution environments and realistic training data. EnvFactory, built by Xu, Wang, and their team (from Huawei and MBZUAI), tackles both automatically.

How It Works

Environment synthesis: EnvFactory autonomously explores and validates stateful, executable tool environments from real resources.
Trajectory synthesis: It produces natural multi-turn trajectories through topology-aware sampling and calibrated refinement — generating queries with *implicit intent*, not "step 1, click here" instruction sequences.

Results

From just 85 validated environments across 7 domains, the pipeline generated 2,575 training trajectories, yielding:

BFCLv3: up to +15% improvement
MCP-Atlas: +8.6%
Conversational benchmarks: +6%

Open Questions

The author flags several unclear points:

1. Environment validation coverage — are 85 environments sufficient to generalize across all domains? 2. Trajectory diversity — if topology-aware sampling converges to a few typical paths, could the RL policy overfit? 3. Safety with real APIs — how does EnvFactory avoid invoking real tools with side effects during testing?

References

1. Xu, M., Wang, Z., Deng, M., et al. (2026). *EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL*. arXiv:2605.18703 [cs.CL]. 2. Mialon, G., et al. (2023). *OpenFunctions: A Large-Scale Evaluation of Tool-Augmented LLMs*. arXiv. 3. Yao, S., et al. (2023). *ToolLLM: Facilitating Large Language Models to Master 16000+ Real-World APIs*. ICLR.