What problem does Model Context Protocol solve for agent systems?

I treat protocol compatibility as only one layer of integration. Server identity, version negotiation, capability changes, timeouts, output schemas, error semantics, provenance, and lifecycle all need application handling. A dynamically discovered tool cannot bypass tenant policy or approval merely because it follows MCP. The host maintains a trust registry and stable internal tool identities so server changes are observable and resumable runs do not unknowingly invoke a different capability.