Teams from Renmin University of China, Fudan University, and Shanghai Jiao Tong University have released ToolSpace, a training environment with 804 real tool environments that lifts Qwen3.6-35B’s average score across seven long-horizon benchmarks from 23.0 to 41.0. We see the key to agent training shifting from “teaching models to guess how tools connect” to “teaching models to learn execution within verifiable business workflows.”
What This Is
ToolSpace is a post-training environment for agents—models that can autonomously plan and call tools to complete tasks. It first collects 200,000 real tool-call traces, then extracts data dependencies, state dependencies, and retry dependencies to map how tools combine. It subsequently generates verifiable training tasks through real calls, failure rollbacks, and result replays.
Rather than asking models to synthesize databases, APIs, and tasks, ToolSpace is closer to the cross-service workflows found in real enterprise systems: checking email, writing to a spreadsheet, and finally sending a notification. The team also builds environments on demand, requiring 55.3% fewer files for 500 environments than Agent-World. The paper, code, and models have all been released.
Industry View
The most striking result is that the 35B model gained 18 points after supervised fine-tuning, then reached 43.2 points after reinforcement learning—where model policies are continually adjusted based on execution outcomes. The 744B version averaged 56.6, approaching the Claude Opus 5 included in the paper’s comparison. These results suggest that whether training data contains real call relationships may improve long-chain execution more than simply increasing parameter count.
But the risks are equally clear. The 200,000 traces come from public tools and services, while the most sensitive data in real enterprises is often restricted by permissions and confidentiality requirements, making those conditions difficult to reproduce. Benchmark gains do not mean immediate deployability: tool definitions change, external APIs fail, and historically successful traces may reinforce outdated workflows. The comparison models and test setup also require independent reproduction. The findings do not justify the claim that smaller models can already match the strongest general-purpose models.
Impact on Regular People
For enterprise IT: Companies can place greater emphasis on call logs, workflow dependencies, and failure samples rather than simply procuring larger models. However, before sensitive business data can be used for training, permission, data masking, and audit issues still need to be resolved.
For individual work: Agents may handle sequential tasks across email, spreadsheets, and collaboration tools more reliably. Critical actions such as approvals, payments, or deletions will still require clear permission boundaries and human review.
For the consumer market: Customer service, shopping, and booking assistants could reduce interrupted journeys and repetitive actions. However, training on real workflows will also increase system maintenance costs, so providers may not make this capability available to everyday users immediately.