ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents

Researchers propose ToolSpace, a tool-centric framework designed to improve long-horizon language model agents by scaling tool-centric environment synthesis. The framework uses deterministic rules to mine dependencies from execution traces of real Model Context Protocol services and open agent corpora, clustering them into environments and generating tasks through state-aware random walks. Experiments show that ToolSpace significantly raises benchmark scores for multiple frontier language models. Open article →