ToolSpace builds agent training worlds from real tool traces

ToolSpace is a method for building training environments for AI agents that start from the tools instead of the data. In a paper submitted to arXiv on November 3, ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents, a team of 25 authors led by Guanting Dong says it lifts the average score of Qwen3.6-35B-A3B across seven tool-use benchmarks from 23.0 to 41.0. Code and model weights are out.

The idea is simple. To train an agent to use tools well, you need many practice worlds with tools, state and tasks. Making those worlds by hand does not scale, so most recent work has an LLM invent them. The ToolSpace authors argue that this is backwards, and they build the worlds from what real tools did.

The problem with starting from data

Existing methods for scaling environments for MCP (Model Context Protocol, the standard way agents connect to outside services) usually begin with a made-up database and scenario. Tools and tasks are then derived from it. The paper names three consequences: the tool interfaces don't match real ones, tool chains were never checked by running them, and tasks rarely cross more than one environment.

Another family of methods uses an LLM as a simulator that pretends to be the tool. The authors note that such a simulator can hallucinate and lose track of state.

What the team did

ToolSpace works in two stages.

Explore the tool space. The team collected more than 200,000 execution trajectories from MCP-related sources such as Toucan-1.5M and ToolBench, plus open agent datasets such as SWE-agent and OpenHands. From them it built a graph of which tool calls follow which. Edges are mined with fixed rules, not an LLM's judgment, and come in three kinds: data (one call's output feeds the next), state (one call changes what the next sees) and retry. The graph has 77,760 call-pattern nodes from 2,042 canonical tools, and about 69% of edges are data edges, 25% state and 6% retry. Clustering it gives 804 environments.

Only environments that change state get a synthesized database. Read-only tools restore snapshots, and stateless tools replay recorded outputs.

Synthesize tasks across environments. The pipeline takes random walks through the graph, choosing the next tool in proportion to how often it followed in real traces, and rolls back when a step fails. Each chain becomes a natural-language task with checks on state, trace and output. A task is kept only if the chain replays successfully from the initial state. The teacher and task-synthesis model was Kimi K3.

The result is 34,376 supervised fine-tuning trajectories and 653 reinforcement learning samples, decontaminated against the test sets.

Results

The team trained three models: ToolSpace-8B on Qwen3-8B, ToolSpace-35B-A3B on Qwen3.6-35B-A3B, and ToolSpace-744B-A40B on GLM-5.2. The 8B and 35B models also have a version with reinforcement learning (GRPO, 48 steps). The 744B model uses supervised fine-tuning only. The paper reports average scores over seven benchmarks: AutomationBench, SkillsBench, JobBench, tau3-Bench Banking, BFCL V4, Harvey LAB and OfficeQA Pro.

On individual benchmarks the 35B model moves from 8.2 to 45.8 on AutomationBench and from 9.3 to 34.0 on tau3-Bench Banking. Gains are not uniform. Harvey LAB goes from 0.8 to 1.7, and OfficeQA Pro from 45.9 to 48.9. The 8B model beats other environment-scaling baselines in the table (EnvScaler-8B, AWM-8B, Agent-World-8B and others, all under 10) by a wide margin.

The authors also built harder "Pro" versions of two benchmarks with their pipeline: 300 AutomationBench Pro tasks and 98 tau3-Bench Banking Pro tasks, reviewed by three experts. Seven frontier models score lower on them, by up to 38.2 points. Claude Opus 5.5 falls from 59.2 to 26.3 on AutomationBench, for example.

Some ablations are worth knowing. Replacing edges seen in real execution with edges judged by an LLM drops the 35B model's AutomationBench score from 45.8 to 37.8. Removing state edges drops its tau3-Bench Banking score from 34.0 to 21.5. Performance rises as data and environments grow.

How far to trust it

The paper is labeled a work in progress, and it has no limitations section. Some caveats come from its own text:

• The cost claim is a proxy. ToolSpace needs 1,218 files against 2,724 for Agent-World on 500 sampled environments, a 55.3% cut, but the paper says this does not measure tokens, time or compute.
• Several baseline scores come from external leaderboards or provider reports, not local reruns.
• The Pro benchmarks report scores for frontier models only, not for the ToolSpace models.
• The headline "matches Claude Opus 5" is an average over seven benchmarks that the same team chose and, for two of them, built a harder variant of. A tie on average hides the spread, as the Harvey LAB score shows.
• A scaling test found that choosing the top half of environments by relevance did no better than a random half, within 1.7 points on four benchmarks. That suggests the gains come more from volume and variety than from picking the right environments.

What is released

The GitHub repository has the task-synthesis pipeline, seven benchmark suites and the evaluation packages for the two Pro benchmarks. The README says the full ToolSpace graph is large, so the release uses AutomationBench as a worked example. It lists 300 AutomationBench Pro tasks, while the paper text describes 600, a mismatch the README does not explain. The README states no top-level license and says third-party components keep their own. Weights for all three models are linked from a Hugging Face collection, and a project page is up.

Full online evaluation needs several outside API keys, including search services and a Hugging Face token. Offline demos and replay do not.

Why it matters

Agents fail on long jobs that chain many tools, and the training data for them is scarce. If environments grounded in real tool behavior train better than invented ones, as these ablations suggest, a small team can close part of the gap to frontier models with open weights and a synthetic pipeline. Independent reproduction, on benchmarks the authors did not build, is what to look for next.