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

ToolSpace. Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents. By Guanting Dong and twenty-four others, from Renmin University of China, Fudan University, and one other institution. Published on the third of November.

Large language models now work as long-horizon agents, composing diverse tools across MCP services into executable workflows. Training them means scaling up MCP environments, and the existing methods for that are data-centric. They synthesize a database and a scenario first, then derive tools and tasks from those. The order has consequences, and the abstract names three. Tool interfaces may not match real ones, so an agent learns to call tools that do not behave like the ones it will meet. Tool compositions are not grounded in execution, so the chains it practises may never have actually run. And tasks rarely span more than one environment, even though the agents themselves compose tools across services.

ToolSpace reverses the order. It is tool-centric: start from how real tools are actually called together, and build environments around that. The dependencies come from execution traces of real MCP services and of open agent corpora, and they are mined with deterministic rules rather than with the judgment of an LLM. The database, which came first in the older recipe, now comes last, and only where it is needed.

From adjacent calls in those traces, ToolSpace extracts three kinds of dependency: data, state, and retry. The result is a weighted tool-composition graph, which is clustered into environments. Databases are synthesized only where an environment mutates state. Query tools and stateless tools simply replay recorded returns. Tasks then come from state-aware random walks over the graph, with rollback. Each walk yields a chain, which is decompiled into a natural-language task with assertion-based checks. A sample is kept only if it replays successfully from the initial state.

The pipeline was first applied to AutomationBench and to tau3-Bench Banking, producing harder Pro versions of each. On those, seven frontier models drop by as much as thirty-eight point two points.

As training data, it was measured across seven long-horizon tool-use benchmarks. Qwen3.6-35B-A3B rises from an average score of twenty-three to forty-one, and to forty-three point two once reinforcement learning is added. GLM-5.2 rises from forty-four point six to fifty-six point six, which puts it on par with Claude Opus 5. Ablations show that edges witnessed in execution clearly outperform edges judged by an LLM, and performance keeps improving as data and environment scale grow.

One choice runs through all of this: where the ground truth comes from. Instead of asking a model which tools belong together, ToolSpace looks at what real executions did, and keeps only tasks that can be replayed. The ablation on execution-witnessed edges is the direct test of that choice, and it comes out in its favour.