EnvFoundry: A Scalable Framework for Tool Environment Synthesis with Controllable Dependency Structures

Large language models (LLMs) are increasingly deployed as agents that solve real-world tasks through tool use, making high-quality interactive environments essential for training and evaluation. Synthesizing these environments involves coordinating three components—task queries, toolsets, and simulators—that are coupled by cyclic constraints: together, they must make tasks solvable while preserving the dependencies that require multi-round interaction. Existing sequential synthesis pipelines often leave these dependencies implicit, requiring later components to accommodate already-fixed artifacts and risking infeasible tasks or unintended shortcuts. We observe that these constraints arise in part from shared relations among entities and their attributes. Based on this insight, we propose EnvFoundry a scalable coarse-to-fine framework that makes this structure explicit through a meta-dependency graph (MDG). The MDG provides coordinated instructions for synthesizing all three components, aligning task requirements, tool capabilities, and environment states while controlling when information becomes accessible. This design supports consistent environment synthesis, controllable task dependencies, and extension to richer tools and state dynamics. Experiments on $\tau$-bench, ${\tau }^2$-Bench, and VitaBench show that the synthesized environments support both supervised fine-tuning and reinforcement learning, substantially improving agentic tool-use performance.