ToolSpace Mines Real MCP Traces to Scale Tool Agents

Researchers at Renmin University of China, Fudan University, and Shanghai Jiao Tong University have released ToolSpace, a data-synthesis pipeline that inverts the standard Model Context Protocol workflow by mining tool dependencies from real execution traces. The method lifts the Qwen3.6-35B-A3B backbone from 23.0 to 41.0 average score across seven long-horizon agent benchmarks, reaching 43.2 after reinforcement learning. Its 744B variant posts 56.6 average, ahead of GPT 5.6 Sol at 48.9 and matching Claude Opus 5 performance, according to the published paper.

How ToolSpace Inverts the MCP Pipeline

Traditional agent data pipelines begin by inventing a database schema, then deriving tools from it. ToolSpace reverses this approach by mining data, state, and retry dependencies from adjacent calls in real MCP execution traces. The system applies deterministic rules to type edges and weights them by how often one tool call actually followed another in live trajectories. No LLM judgment is used in the reconstruction process, the authors emphasize.

The resulting weighted graph contains 77,760 tool-call-pattern nodes spanning 2,042 canonical tools, organized into 804 real MCP environments. These environments cover Gmail, Google Sheets, Slack, Google Drive, Salesforce, and Airtable. Data edges account for 69 percent of the graph, state dependencies for 25 percent, and retry patterns for 6 percent.

Frontier Models Drop on Harder Benchmarks

The research team applied the same mining machinery to two existing benchmarks, producing AutomationBench Pro with 300 tasks and τ³-Bench Banking Pro with 98 tasks. Three doctoral-level experts reviewed each benchmark over a two-day period, achieving Fleiss κ=0.82 and 93.1 percent observed agreement. Seven frontier models drop by up to 38.2 points on AutomationBench Pro compared to the original split. Claude Opus 5.5 falls 32.9 points, while GPT 6.1 Sol drops 26.1. On τ³-Bench Banking, the declines land at 15.8 and 16.8 respectively.

Average tool calls per task rise from 26.8 to 43.6 under Standard conditions to 56.3 to 60.6 under Pro conditions on AutomationBench, confirming that the added difficulty stems from longer and more frequent tool orchestration, the paper states.

Release Date and Training Data

ToolSpace was officially released on November 4, 2026. The authors published the complete codebase, model weights, and 34,376 supervised fine-tuning samples plus 653 reinforcement learning samples generated by the pipeline. The release arrives during a busy season for agent evaluation, with 526 prior agent stories logged over the past 90 days across public tracking streams.

Performance Across Model Sizes

The 35B-parameter Qwen3.6-35B-A3B variant demonstrates a 78 percent improvement over its baseline, rising from 23.0 to 41.0 average score across the seven benchmarks. After reinforcement learning, the score climbs to 43.2. The larger 744B variant achieves 56.6 average, placing it ahead of GPT 5.6 Sol and level with Claude Opus 5 on the tested tasks. These results suggest that mining real tool-use patterns from production MCP environments produces training data that generalizes better to complex orchestration scenarios than synthetic data pipelines.

What This Means

ToolSpace represents a shift from synthetic data generation to trace-based reconstruction for training long-horizon tool agents. By grounding the data pipeline in actual MCP execution patterns rather than hypothetical database schemas, the method captures real-world tool dependencies that synthetic approaches often miss. The sharp performance drops among frontier models on the new Pro benchmarks indicate that current evaluation sets may underestimate the difficulty of multi-step tool orchestration in production environments. As agent deployments expand into workflows with dozens of sequential tool calls, pipelines that mine real dependencies may prove more effective than those that invent them.