ToolSpace scales tool-centric environment synthesis for long-horizon agents

ToolSpace raises Qwen3.6-35B-A3B from 23.0 to 41.0 average score and GLM-5.2 from 44.6 to 56.6 on seven long-horizon tool-use benchmarks.

Large language models serve as long-horizon agents composing diverse tools across MCP services into executable workflows. ToolSpace is a tool-centric framework that mines data state and retry dependencies from adjacent calls in execution traces of real MCP services and open agent corpora using deterministic rules and clusters the resulting weighted tool-composition graph into environments. Databases are synthesized only for state-mutating environments while query and stateless tools replay recorded returns. It performs state-aware random walks with rollback and decompiles each chain into a natural-language task with assertion-based checks. Applied to AutomationBench and tau3-Bench Banking the pipeline produces harder Pro versions on which frontier models drop up to 38.2 points. Across seven long-horizon tool-use benchmarks ToolSpace raises Qwen3.6-35B-A3B from 23.0 to 41.0 average score and GLM-5.2 from 44.6 to 56.6.