Researchers test role-based tool access for AI agents

Researchers have proposed a framework that limits which tools an AI agent can use as it works. Their preprint describes skilder, which gives an agent a small catalog of roles and supplies the tools for a role only after the agent selects it. A server then enforces the limits attached to those tools. Across 13 tasks tested with six models, the simulated authorization layer executed no unauthorized tool calls or parameter violations when agents completed discovery and made a governed call. Some agents still failed to follow the discovery process or meet response-quality checks, so the result does not establish reliable task completion.