ToolSpace builds agent training environments from real MCP tool traces
ToolSpace mines tool dependencies from real MCP execution traces, turns them into environments and checked tasks, and reports large gains for open models on long-horizon tool use.
The problem with data-centric environments
Large language models are deployed as long-horizon agents that compose diverse tools across MCP services into executable workflows. According to the ToolSpace paper (arXiv 2611.05036, 25 authors), existing MCP environment-scaling methods are data-centric: they first synthesize a database and a scenario, then derive tools and tasks from them.
The authors list three consequences of that order. Tool interfaces may not match real ones, tool compositions are not grounded in execution, and tasks rarely span more than one environment.
Reversing the order
ToolSpace is a tool-centric framework that reverses this order, so the tools come first. It mines data, state and retry dependencies from adjacent calls in execution traces of real MCP services and open agent corpora. The abstract stresses that this step uses deterministic rules rather than LLM judgment. The result is a weighted tool-composition graph, which ToolSpace clusters into environments.
Databases are synthesized only for state-mutating environments, while query and stateless tools replay recorded returns. To produce tasks, ToolSpace performs state-aware random walks with rollback and decompiles each chain into a natural-language task with assertion-based checks. Only samples that replay successfully from the initial state are kept.
Results as reported
All figures are the authors’ own, from the abstract.
- Applied to AutomationBench and tau3-Bench Banking, the pipeline produces harder Pro versions, on which 7 frontier models drop by 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, or 43.2 with RL.
- On the same benchmarks it raises GLM-5.2 from 44.6 to 56.6, which the authors put on par with Claude Opus 5.
- Ablations show that execution-witnessed edges clearly outperform LLM-judged ones, and performance keeps improving with data and environment scale.
The ablation result is the core argument of the paper: dependencies taken from real executions make better training material than tool links that a language model judged plausible.
What the abstract leaves open
The abstract does not say how many environments or tasks ToolSpace generated, which MCP services and open agent corpora the traces came from, or which seven benchmarks make up the reported average. It does not name the 7 frontier models or give their scores before the drop on the Pro versions. The paper lists its code repository: ToolSpace on GitHub.
Source: ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents (arXiv 2611.05036)