EnvFactory: 85 Simulated Environments Train a Generalist 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 things: scalable executable environments and realistic training data. EnvFactory, from a team at Huawei and MBZUAI (Xu, Wang, et al.), addresses both automatically. The framework autonomously explores and verifies stateful, executable tool environments built 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-by-step instructions. Using only 85 verified environments spanning 7 domains, it produced 2,575 training trajectories, yielding gains of up to +15% on BFCLv3, +8.6% on MCP-Atlas, and +6% on conversational benchmarks. Open questions include whether 85 environments generalize across domains, whether topology sampling converges to a few typical paths and causes RL policy overfitting, and how EnvFactory ensures safety when interacting with real APIs that may have side effects.

Agentic RL training for tool-use capability faces two bottlenecks: scalable execution environments and realistic training data. EnvFactory, developed by Xu, Wang, and colleagues (Huawei, MBZUAI), tackles both automatically.

Approach

Environment synthesis: EnvFactory autonomously explores and verifies stateful, executable tool environments derived from real resources.
Trajectory generation: Natural multi-turn trajectories are synthesized through topology-aware sampling and calibrated refinement — producing queries with implicit intent rather than "click here, step 1" instruction sequences.
Scale: Only 85 verified environments across 7 domains generated 2,575 training trajectories.

Results

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

Open Questions

Environment verification coverage: Are 85 environments sufficient to generalize across all domains?
Synthetic trajectory diversity: If topology-aware sampling converges to a few typical paths, could the RL policy overfit them?
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.