LDP: An Identity-Aware Protocol for Multi-Agent LLM Systems

As multi-agent AI systems scale to production, the protocols connecting them constrain their capabilities. Current protocols such as Google's A2A and Anthropic's MCP omit model-level metadata, preventing quality-aware routing, efficient communication, and governance. We present the LLM Delegate Protocol (LDP), which introduces five AI-native primitives: rich delegate identity cards, progressive payload negotiation, governed sessions, structured provenance, and trust domains. We implement LDP as a plugin adapter for the JamJet agent runtime and evaluate it against A2A and random baselines. Identity-aware routing achieves 12x lower latency on easy tasks; semantic frame payloads reduce token count by 37 percent with no quality loss; governed sessions eliminate 39 percent token overhead; and noisy provenance degrades quality below the no-provenance baseline, showing that confidence metadata is harmful without verification. Trust domain and fallback analyses demonstrate architectural advantages in security and resilience. This paper contributes a protocol design, reference implementation, and initial evidence that AI-native protocol primitives enable more efficient and governable multi-agent delegation.