ToolSpace Synthesizes MCP Environments for Long-Horizon Agents
Hugging Face · Papers reports that Guanting Dong and 24 coauthors proposed ToolSpace, a framework that constructs tool-centric environments from execution traces of real MCP services and open agent corpora. It uses deterministic rules to derive tool dependencies, state-aware random walks with rollback, and assertion-based checks to retain tasks that replay successfully. The paper reports that seven frontier models dropped by as much as 38.2 points on harder AutomationBench and tau3-Bench Banking Pro versions, while Qwen3.6-35B-A3B and GLM-5.2 improved across seven long-horizon tool-use benchmarks.
ToolSpace offers a concrete execution-grounded approach to scaling multi-environment MCP tasks, and its reported benchmark gains make the framework relevant to researchers building and evaluating long-horizon agents.
Computer Science > Computation and Language
arXiv:2611.05036 (cs)
[Submitted on 3 Nov 2026]
Title:ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents
Authors:Guanting Dong, Zhichao Sun, Xinyu Yang, Rui Wang, Xin Deng, Jiajie Jin, Zhixiong Yang, Junhao Huang, Han Han, Mingyu Ma, Xiaotao Wang, Yuyang Hu, Xiaoshuai Song, Xiaoxi Li, Jiejun Tan, Zhecan Wang, Shihan Dou, Honglin Guo, Zhiheng Xi, Tao Gui, Tao Ji, Qi Zhang, Yutao Zhu, Ji-Rong Wen, Zhicheng Dou
View a PDF of the paper titled ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents, by Guanting Dong and 24 other authors
#ToolSpace#AutomationBench#tau3-Bench Banking#Qwen3.6-35B-A3B#GLM-5.2#Claude Opus 5