Tool permissions are access-control policies enforced by the agent runtime or harness that determine which tools an agent is allowed to call, with what arguments, and under what conditions. Unlike model-level instruction following (which is probabilistic), tool permissions are hard guardrails — the runtime will reject unauthorized calls regardless of what the model requests. Permissions may be scoped by agent identity, task type, user role, or risk level, and are a primary mechanism for sandboxing agent behavior in production.