ToolSpace — environments built from observed tool compositions

TL;DR

ToolSpace constructs training environments from dependencies observed in tool execution traces, then keeps tasks that replay successfully. Authors report improved long-horizon tool-use scores; these are preprint results, not independent replication. (source)

Authors & Org

Guanting Dong and 24 coauthors; the complete author list is in the captured arXiv record. Affiliations are not established by the reviewed abstract page. First submitted 2026-11-03. (source)

Method

• Mine data, state and retry dependencies from adjacent calls in real MCP-service traces and open agent corpora using deterministic rules; cluster the weighted composition graph into environments. (source)
• Synthesize databases for state-mutating environments, while query and stateless tools replay recorded returns. State-aware random walks with rollback generate chains, which become natural-language tasks with assertion-based checks. Retain only tasks that replay from the initial state. (source)

Results

• Harder Pro variants of AutomationBench and tau3-Bench Banking reduce scores of 7 frontier models by up to 38.2 points. (source)
• Across seven long-horizon tool-use benchmarks, Qwen3.6-35B-A3B rises from 23.0 to 41.0 average score, or 43.2 with RL; GLM-5.2 rises from 44.6 to 56.6, described as matching Claude Opus 5 in this evaluation. (source)
• Authors report that execution-witnessed graph edges outperform LLM-judged edges, with further improvements as data and environment scale increase. (source)

Significance

For MCP — Model Context Protocol and Agents (LLM Agents), the contribution is grounding synthetic workflows in observed compositions and executable checks. Interpretation: realistic interfaces and replayable tasks are distinct from merely generating plausible scenarios. (source)

Open Questions

The abstract does not quantify total construction cost or demonstrate that replayed query returns reproduce changing live services. Full-paper methods, code and independent reproduction remain unreviewed here. (source)

Cite

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

arXiv:2611.05036, first submitted 2026-11-03. (source)