Skilder Uses Progressive Skill Discovery to Govern Tool-Using LLM Agents

Summary

The paper introduces skilder, a framework for governing large language model agents that use enterprise tools. It addresses the risks of exposing an agent to a large flat toolset, including oversized contexts, poorer tool selection, and the weakness of relying on prompt-defined policies as probabilistic guidance. Skilder packages tools, instructions, skills, and their limits into roles. An agent starts with a minimal role catalog, discovers which roles a task requires, and receives the corresponding capabilities through a single Model Context Protocol server. Tools are available only through learned skills, allowing the server to enforce the learned scope deterministically. In an evaluation covering 13 tasks, six models, and 10 runs per model, the framework’s simulated authorization layer blocked unauthorized tool calls and parameter violations, including spending-limit breaches, whenever models completed discovery and made governed calls. Overall task pass rates also reflected failures to follow the discovery protocol or response-quality checks; the authors distinguish these from authorization failures. The framework further allows agents to acquire capabilities from additional roles during a task, preserving cross-role problem-solving flexibility while retaining system-level governance boundaries.