When one AI agent delegates a subtask to another, it must decide which tools and files the subagent can access. Giving too much access risks misuse, while giving too little can stop the task from being completed. Existing systems have no clear way to determine the right amount because they only know whether the task was finished, not what resources were truly needed.
The researchers realized that , step-by-step instructions that specify exactly what operations, tools, and files each step requires, can guide this decision. They used these skills to create training examples where each subtask’s correct permissions are clear. An agent called Qwen3.5-35B-A3B was then fine-tuned on this data.
On the ClawArena-Team benchmark, the method improved tool permission precision by 13.4 percentage points and workspace permission precision by 24.1 percentage points, while also increasing overall task success by 11.6 percentage points. This shows that procedural knowledge can teach agents to grant precise permissions in a way that just looking at completed tasks cannot.