EnvFactory: When AI Builds Its Own Arsenal for Tool-Use Agent Training

Forum topic · 小凯 · 2026-05-25

Summary

Training AI agents to use tools via reinforcement learning has long been bottlenecked by the lack of executable training environments: real APIs are costly and fragile, while LLM-generated simulated environments are unreliable and lose state. Even when workable environments exist, prior training tasks use overly explicit instructions (e.g., "click the button, type apple"), leaving agents unable to handle implicit human intent. EnvFactory, released May 2026, addresses both problems with a fully automated framework that explores real-world resources, verifies them, and synthesizes stateful, executable tool environments without human intervention, then generates high-quality, multi-turn training data via topology-aware sampling. Using only 85 verified environments and roughly 2,000 synthesized data samples, the framework trained Qwen3 models to improve performance on the BFCLv3 tool-calling benchmark by 15%, outperforming prior methods that used 5x more environments. The paper, arXiv:2605.18703, is titled "EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL."

Introduction: The Embarrassment of Armchair Training

Teaching AI to use tools is like teaching a person to ride a bicycle — reading the manual is useless without a real bike and a practice ground. In academia this is called Agentic Reinforcement Learning (RL). But where do you find enough training grounds? Calling real external software APIs is too expensive and risky. Asking an LLM to "build" a simulated environment often produces nonsense, and saved state from yesterday may vanish today.

Without a real training ground, AI agents can only practice on paper — the skills they learn are superficial.

The Problem: Fake, Mechanical "Textbooks"

Even when a barely usable simulated environment exists, the training material is poor.

Previously constructed practice tasks used overly rigid instructions, such as "click the button in the upper-left corner to open the menu, then type apple." But in the real world, users only say things like "I want to buy two pounds of Fuji apples." As a result, agents trained on such data are used to spoon-feeding and go blank when user intent becomes implicit.

> 💡 Tip: This is called the "Lack of Implicit Human Reasoning." In short, training data looks too much like machine-generated code, with no human flavor, so the AI cannot generalize.

The Solution: EnvFactory's Creation Power

In May 2026, EnvFactory emerged. Since there were no training grounds or good textbooks, it built its own "arsenal":

1. Automated mining and factory building: It searches the web for real resources, automatically explores and builds genuinely executable tool environments with "state memory." Without human intervention, the environments it builds can withstand verification. 2. Topology-based sampling for textbooks: With a training ground, you still need exam questions. EnvFactory senses the environment's topological structure and automatically generates high-quality practical data with multi-turn dialogue and implicit intent.

This creation logic can be captured in a single self-sustaining formula:

$ \mathcal{T}_{data} = \text{Refine}\left( \text{Sample}_{\text{topology}}(\text{Verify}(\text{Explore}(R_{real}))) \right) $

> 💡 Formula notes: It first explores and verifies executable environments from real resources ( $R_{real}$ ), then samples and refines based on the environment's structural topology, ultimately producing high-quality training data ( $\mathcal{T}_{data}$ ).

Here's how powerful this "arsenal" is:

Results: Winning Against the Odds

The results were striking.

Researchers had it build only 85 verified environments, producing just over 2,000 data samples. Training the Qwen3 model with this small dataset caused scores on BFCLv3, a top-tier tool-calling benchmark, to surge by 15% — crushing prior methods that used 5x more environments.

With EnvFactory, AI has finally achieved "training freedom." It can build its own targets, shoot its own arrows, and review its own performance. Self-sufficient, needing nothing external.

---

Paper Reference

All core claims in this article come from the following source.

Paper title: EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL
Release date: May 20, 2026
Identifier: arXiv:2605.18703
Core contribution: Solving the bottlenecks of scarce executable environments and training data lacking real human intent in agentic reinforcement learning.
Mechanism: A fully automated framework, EnvFactory, enabling autonomous exploration and verification of stateful tool environments, plus topology-aware synthesis of high-quality practical training data.