ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents
AuthorsGuanting 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
AffiliationsRenmin University of China Fudan University Shanghai Jiao Tong University
ToolSpace builds tougher, more realistic tool-use environments from real execution traces to train and evaluate long-horizon AI agents.
Key results
Nodes in the execution-grounded tool-composition graph.
Real environments represented in the ToolSpace inventory.
Total trajectories generated by the ToolSpace pipeline.
Seven-benchmark average after ToolSpace training, up from 23.0.
Point decrease for a frontier model compared with the standard benchmark.
Fewer files synthesized than with the compared environment-generation approach.
What the paper found
ToolSpace changes how long-horizon agent environments are built: instead of inventing databases first and deriving tools from them, it mines real tool-use traces to create a weighted graph of executable tool relationships. Deterministic rules identify data, state, and retry dependencies between adjacent calls; state-aware walks then sample tool chains, roll back failed steps, and keep only tasks that pass replay and assertion-based checks. Databases are synthesized only for tools that mutate state, while query and stateless tools can reuse recorded returns. The resulting graph contains 77.76K tool-call patterns, drawn from 2,042 tools across 804 MCP environments, and the pipeline produces 34.376K trajectories. On seven benchmarks, Qwen3.6-35B-A3B improves from 23.0 average score to 41.0 with ToolSpace training, or 43.2 with reinforcement learning; the 744B ToolSpace model scores 56.6, matching Anthropic’s Claude Opus 5 and exceeding OpenAI’s GPT 5.6 Sol at 48.9. The evaluations also include GLM, Meta’s Muse-Glimmer-30B, and DeepSeek models. On harder Pro versions of AutomationBench and τ³-Bench Banking, the largest frontier-model drop is 38.2 points. Selective environment construction also reduces synthesized files by 55.3%. Ablations find that execution-witnessed edges outperform LLM-judged connections, supporting the paper’s central claim: scalable agent training depends on grounded tool composition, not just larger simulated environments.
Original abstract
Large language models are deployed as long-horizon agents that compose diverse tools across MCP services into executable workflows. Existing MCP environment-scaling methods are data-centric: they first synthesize a database and scenario, then derive tools and tasks from them. As a result, tool interfaces may not match real ones, tool compositions are not grounded in execution, and tasks rarely span more than one environment. We propose ToolSpace, a tool-centric framework that reverses this order. ToolSpace mines data, state, and retry dependencies from adjacent calls in execution traces of real MCP services and open agent corpora, using deterministic rules rather than LLM judgment, 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. ToolSpace then performs state-aware random walks with rollback and decompiles each chain into a natural-language task with assertion-based checks, keeping only samples that replay successfully from the initial state. 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 (43.2 with RL) and GLM-5.2 from 44.6 to 56.6, 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.
Read the original paperMore in AI Agents
Browse all 71 papers →Route Anything: A Study of Model and Policy Routing for General-Purpose AI Systems
Pengfei Zhou, Yixing Ma, Hexin Wang, Weidong Tang, Zhenglin Wan, Wangbo Zhao, Yang You
Route Anything studies how AI systems can choose the right specialized model or policy for each task to improve results while spending less compute.
What Effort Level Should I Choose? On the Economics of Reasoning Effort in LLM Agents
Kaiser Sun, Yongchan Kwon, Zain Hasan, Shang Zhu, Kaitlyn Zhou, Federico Bianchi
More reasoning effort can cost much more without reliably improving an LLM agent’s result, so escalating only after verified failure may be a better strategy.
Build, Don't Play: Self-Improving LLM Research Agents Develop a Hearthstone Battlegrounds AI That Rivals Top Players
Changlian Ma, Limin Wang
A team of LLM research agents builds a Hearthstone Battlegrounds AI that reportedly outperforms top human players on the live ladder.