Automatically generated environments have become a scalable way to produce interaction data for training general-purpose agents, but most existing methods generate tasks inside a single environment. Real-world workflows, by contrast, require agents to carry information and actions across multiple services. A new research paper, CompoWorld: Compositional Environment Scaling for General Agents, proposes expanding the task space by composing a finite library of reusable services.
In the proposed pipeline, coding agents convert tool specifications into verified services with typed states and shared interfaces, while a world model handles tools that cannot be reliably implemented. A random-walk procedure then connects services through dependency graphs, allowing tasks to be generated and verified when information must flow from one service to another.
Training relies on two stages. Verified trajectories feed supervised fine-tuning (SFT), while a Completion-Focused Rubric Reward steers reinforcement learning (RL) toward full task completion by emphasizing criteria with lower pass rates within each rollout group.
The authors report constructing 448 services exposing 10,130 tools, and using 3K SFT trajectories and 1K RL tasks to train Qwen3.6-35B-A3B. According to the paper, CompoWorld improves on its backbone by 9.17 points on average across eight benchmarks. On AutomationBench, it is said to surpass frontier models such as Claude Opus 4.6 and to lead all compared agent-specialized 35B-A3B models.