Wenkai Yu, Ning Ding, Wentao Zhang and Yehui Tang submitted a preprint on 6 November 2026 proposing a way to learn subagent authorization from Agent Skills. The work addresses a delegation problem: deciding which tools and workspace resources a subagent needs alongside the work it should perform.[1]
The authors argue that task outcomes and execution histories alone provide limited supervision for judging appropriate permissions. Agent Skills describe operations, tools and files step by step. Their method synthesizes tasks and procedures from those Skills, executes them through subagent delegation in interactive environments, and uses trajectories described as verified to construct execution and authorization training data. The saved abstract does not explain how that verification works.[1]
For Qwen3.5-35B-A3B on ClawArena-Team, the abstract reports gains of 13.4 percentage points in tool permission precision, 24.1 percentage points in workspace permission precision and 11.6 percentage points in overall task success. It also describes the method as competitive with existing agent data synthesis approaches on Terminal-Bench 2.1 and SkillsBench. These are the authors’ reported results; the saved page does not supply the comparison details or permission metric definitions needed to assess them independently.[1]
[1]: From Procedures to Permissions: Learning Precise Subagent Authorization from Agent Skills, arXiv, 6 November 2026.
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
814b16b43a211b0cb6882b1108c5fbdc3b91a7c2ea1def1274d4ec0a088b37f1