EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL
Source: HuggingFace Daily Papers 2026-05-20 · arXiv 2605.18703 · raw
TL;DR
Equipping LLMs with tool-use via Agentic RL is bottlenecked by two problems: lack of scalable robust execution environments, and scarcity of realistic training data that captures implicit human reasoning. Existing approaches use costly real-world APIs, hallucination-prone LLM simulators, or synthetic environments that are single-turn or depend on pre-collected documents. Synthetic trajectories are often over-specified, resembling instruction sequences rather than natural human intents, which reduces their RL training value. EnvFactory autonomously explores and verifies stateful executable tool environments from authentic resources, and synthesizes natural multi-turn trajectories through topology-aware sampling and calibrated refinement to produce grounded queries with implicit intents. Using only 85 verified environments across 7 domains, EnvFactory generates 2,575 SFT and RL trajectories. Despite using 5x fewer environments than prior work, it produces +15% on BFCLv3, +8.6% on MCP-Atlas, and +6% on conversational benchmarks including tau^2-Bench and VitaBench.
Why it matters
The agentic RL stack has been bottlenecked by environment cost. Real APIs are expensive and brittle. LLM simulators hallucinate. Static synthetic environments don't generalize. EnvFactory's two-step pipeline (autonomous environment verification, then topology-aware trajectory generation) compresses both costs into a single automated framework, and the 5x environment efficiency at +15% BFCLv3 is large enough to indicate the bottleneck has shifted. The implicit-intent observation is the substantive insight: prior synthetic trajectories read like instruction sequences, real human intents are underspecified, and the gap is what makes RL training transfer.
Mechanism
Two pipelines. (1) Environment synthesis: start from authentic resources (docs, schemas, API specs); autonomous exploration generates candidate stateful environments; verification checks the environment is executable and well-formed. (2) Trajectory generation: topology-aware sampling traverses the environment graph in human-plausible ways; calibrated refinement removes the instruction-sequence over-specification and leaves implicit intents.
Open questions and gaps
The "authentic resources" prerequisite is the limiting factor. EnvFactory works from existing docs and schemas; whether it can produce environments for tasks where the resource is sparse or proprietary is open. The 85 environments span 7 domains, which is wide for a paper but narrow versus the universe of real production environments. The implicit-intent calibration is the load-bearing module and the paper presumably gives details that are not in the abstract.
Industrial implication
Production agentic RL is gated by environment construction. A pipeline that turns docs into verified stateful environments with realistic trajectories is exactly what frontier labs and serious agentic startups need. If EnvFactory's verification is robust across domains beyond the 7 tested, it becomes a default substrate for agentic RL within 90 days.
Connections
- OpenComputer (today, 2605.19769) builds verifier-grounded software worlds for computer-use agents. The two papers attack the same problem (verifiable environments for agent training) from different scales: EnvFactory builds many small environments from docs, OpenComputer builds a deep desktop-application substrate with hard-coded verifiers.
- AgentKernelArena (2026-05-19) measures the seen-versus-unseen-shape generalization gap in GPU-kernel agents. EnvFactory's topology-aware sampling is one possible answer to whether broader environment coverage closes that gap.