TRUSS Ensures Safe, Reliable AI Agent Skill Generation

Automatically generating skills for AI agents can significantly enhance their capabilities, but ensuring these skills are both effective and safe is a major challenge. Traditional evaluation methods often fall short, failing to fully assess the actions an agent will take or the potential side effects. This paper introduces TRUSS, an evidence-guided framework designed to generate AI agent skills that are task-reliable and user-safe. TRUSS employs a two-stage validation process: first, it statically inspects functional claims against source evidence and evaluates the skill artifact against nine predefined safety properties. Skills that pass this initial gate are then loaded into a "shadow agent" within a controllable execution environment. Here, brokered tools expose requested actions to policy enforcement, and execution traces are recorded to preserve provenance. Any functional failures or safety violations are linked back to the responsible skill content, guiding an iterative refinement process. TRUSS significantly improves task effectiveness and achieves 100% security rates in benchmarks, demonstrating its ability to generate robust and safe agent skills.