ToolSpace mines real MCP execution traces to synthesize tool environments for long-horizon agents
RUC · hf · 2026-11-04
Researchers from Renmin University, Fudan and Shanghai Jiao Tong propose ToolSpace, a tool-centric framework for synthesizing environments for long-horizon agents.
- Reversed pipeline: existing MCP environment-scaling methods are data-centric (synthesize databases first, derive tools later), producing unrealistic tool interfaces; ToolSpace mines data, state and retry dependencies from execution traces of real MCP services and open agent corpora, using deterministic rules instead of LLM judgment, then clusters weighted tool-composition graphs into environments
- Databases are synthesized only for state-mutating environments; query/stateless tools replay recorded returns
- Tasks are generated via state-aware random walks with rollback, decompiled into natural-language tasks with assertion-based checks, keeping only verifiable samples
A more execution-grounded way to evaluate and train tool-using agents.
More from coding & agent
- EvoTeam evolves multi-agent team organization and skills from execution experience — JLU-Deep · 2026-11-04
- Switch-SWE benchmark shows coding agents lose focus when switching unfinished tasks — SJTU · 2026-11-04
- Fan builds draggable "Tibo calibrator" meme page with Nano Banana and RIFE for $0.24 — GlobalProcedure1731 · 2026-11-04
- Valid tokens, allowed actions, zero intent checks: the agent security gap nobody solves — Cytersmyz · 2026-11-04
- Stanford CS329X HW3: Coco agent traces plus AutoMetrics evaluators, open-sourced — EchoShao8899 · 2026-11-04
- Netflix shares agentic automation for auto-diagnosing failed data and ML workloads — AxSaucedo · 2026-11-04