From Procedures to Permissions: Learning Precise Subagent Authorization from Agent Skills
Organizations: Peking University · Samsung Research, Beijing, China
Abstract
As language agents tackle increasingly complex tasks, they increasingly rely on multiple agents to coordinate specialized work, requiring decisions not only about what each subagent should do but also which tools and workspace resources it should access. Existing agent systems provide mechanisms for controlling tools and execution environments, but do not directly specify the resources required by each delegated subtask. Moreover, task outcomes and execution trajectories alone provide limited supervision for determining whether a subagent has received appropriate permissions. We observe that Agent Skills encode procedural knowledge describing the operations, tools, and files required at each step, providing a natural source of supervision for authorization. Building on this observation, we synthesize tasks and step-by-step procedures from Skills, execute them through subagent delegation in interactive environments, and use verified trajectories to construct training data for task execution and authorization. On ClawArena-Team, our method improves tool permission precision by 13.4 percentage points and workspace permission precision by 24.1 percentage points, while increasing overall task success by 11.6 percentage points for Qwen3.5-35B-A3B. On Terminal-Bench 2.1 and SkillsBench, our method remains competitive with existing agent data synthesis approaches. These results show that the procedural knowledge embedded in Agent Skills provides a source of supervision for precise subagent orchestration that task outcomes alone cannot provide.
Figures & tables
| Method | SMS | TCR | TPP | ROC | WPP | MCA |
| Ours | 44.6 | 56.6 | 73.3 | 98.1 | 59.5 | 84.5 |
| Base | 30.8 | 45.0 | 59.9 | 96.1 | 35.4 | 82.8 |
| w/o procedure | 37.5 | 53.5 | 63.6 | 94.7 | 38.3 | 84.0 |
| w/o verification | 32.8 | 49.2 | 57.7 | 96.9 | 30.5 | 81.9 |
| w/o authorization supervision | 24.5 | 40.7 | 50.9 | 88.8 | 27.0 | 73.7 |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| Configuration | Value |
| GPU model | H800 |
| GPU count for training | 8 (Qwen3.5-4B); 24 (Qwen3.5-35B-A3B) |
| Training epochs | 10 |
| Batch size | 256 |
| Loss normalization | per-token loss |
| Data order | shuffled |
LLM Agent Orchestration
Large language model agent orchestration coordinates models, tools, and workflows to complete complex tasks reliably and efficiently.
See all 494 papers