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

📝 Paper Summary

Multi-agent communication protocols Agent interoperability Identity-aware routing

LDP is an AI-native communication protocol that exposes model-specific properties (identity, cost, quality) to enable efficient delegation, structured negotiation, and governance in multi-agent systems.

Core Problem

Current agent protocols like A2A and MCP treat models as opaque services, hiding critical metadata (parameter count, reasoning capability, cost) needed for effective delegation and routing.

Why it matters:

Routers cannot distinguish between a cheap 3B model and a reasoning-heavy 8B model based solely on generic skill names, leading to inefficient assignments
Stateless protocols require re-transmitting full conversation history for every request, incurring quadratic token costs
Lack of provenance tracking makes it impossible to verify the reliability or confidence of synthesized outputs from multiple agents

Concrete Example: When a router must choose between a 3B classification model and an 8B reasoning model, A2A only shows 'classification' vs 'reasoning' skills. The router cannot assess cost-quality tradeoffs, leading to using an expensive model for a simple task or a weak model for a complex one.

Key Novelty

AI-Native Delegate Protocol (LDP)

Introduces 'Delegate Identity Cards' that expose deep model properties (model family, reasoning profile, cost/quality hints) rather than just skill lists
Implements 'Progressive Payload Modes' where agents negotiate the most efficient communication format (e.g., structured JSON vs. text), falling back if validation fails
Establishes 'Governed Sessions' that maintain context state server-side to eliminate the token overhead of stateless re-invocation

Evaluation Highlights

Identity-aware routing achieves ~12x lower latency on easy tasks compared to baselines by correctly dispatching to smaller specialized models
Semantic frame payloads reduce token count by 37% (p=0.031) compared to natural language without observed quality loss
Governed sessions eliminate 39% token overhead at 10 conversation rounds compared to stateless re-invocation

Breakthrough Assessment

7/10

Strong architectural contribution addressing the specific needs of LLM agents (token costs, context limits) rather than generic API calls. Empirical results on token savings are significant, though the delegate pool size in experiments was small.

⚙️ Technical Details

Problem Definition

Setting: Multi-agent delegation where a router/caller must select a delegate agent and exchange messages to solve a task

Inputs: Task prompt, set of available delegate agents with metadata

Outputs: Task result with provenance metadata

Pipeline Flow

Discovery (Reading Delegate Identity Cards)
Session Negotiation (Handshake & Capability Manifest)
Task Delegation (Payload Transmission)
Result Synthesis (Provenance Tracking)

System Modules

Identity Card

Expose model properties including family, parameter count, quality hints, and cost profiles

Model or implementation: Metadata Schema

Session Manager

Negotiate payload modes and maintain session state (budget, privacy context)

Model or implementation: Protocol State Machine

Policy Engine

Validate tasks against configurable rules (capability scope, jurisdiction, trust domains)

Model or implementation: Rule-based Validator

Novel Architectural Elements

Progressive Payload Modes: A tiered communication standard (Text, Semantic Frames, Embeddings, etc.) with automatic negotiation and fallback
Delegate Identity Card: A specialized metadata structure for AI models (weights fingerprint, reasoning profile) integrated into the discovery layer
Trust Domains: Protocol-level security boundaries that enforce policy beyond simple bearer token authentication

Modeling

Base Model: JamJet Runtime (Rust) + Ollama (Local Models)

📊 Experiments & Results

Evaluation Setup

Routing and delegation tasks using local Ollama models, evaluated by Google Gemini 2.5 Flash as judge

Benchmarks:

Routing Quality Task Set (30 tasks (easy/medium/hard) across classification, reasoning, coding, math) [New]
Payload Efficiency Task Set (20 tasks for reasoning handoffs, context transfer, verification) [New]
Security Simulation (100 security injection scenarios) [New]

Metrics:

Latency
Token Count
Aggregate Quality Score (1-10)
Attack Detection Rate
Statistical methodology: Mann-Whitney U test for significance; Cohen's d for effect size

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
Easy Tasks (Routing)	Latency	Not reported in the paper	Not reported in the paper	Not reported in the paper
Payload Efficiency Tasks	Token Count Reduction	Not reported in the paper	Not reported in the paper	-37%
Multi-round Delegation (10 rounds)	Token Overhead	Not reported in the paper	Not reported in the paper	-39%
Attack Detection	Detection Rate	6%	96%	+90%
Failure Recovery	Task Completion Rate	35%	100%	+65%

Main Takeaways

Exposing model identity enables massive latency reductions (~12x) on easy tasks by routing them to smaller, faster models instead of generic large ones
Structured payloads (Semantic Frames) significantly reduce token costs (-37%) without degrading task quality
Stateful sessions are critical for multi-turn interactions, reducing token overhead by ~40% compared to stateless baselines like A2A
Noisy provenance metadata can actually harm synthesis quality if not verified, performing worse than no provenance at all

📚 Prerequisite Knowledge

Prerequisites

Basic understanding of LLM agent architectures (routing, tool use)
Familiarity with API protocols (REST, RPC)
Knowledge of token-based cost models in LLMs

Key Terms

A2A: Google's Agent-to-Agent Protocol—a service-oriented interface where agents expose skills but lack model-specific metadata

MCP: Model Context Protocol—Anthropic's standard for connecting AI assistants to external tools and data sources

semantic frames: Mode 1 payload in LDP; a typed structured JSON format with explicit fields (task_type, instruction) to reduce verbosity compared to natural language

provenance: Metadata tracking the origin, confidence, and verification status of an AI-generated result

trust domain: A security boundary within which identity, policy, and transport guarantees are recognized and enforced

governed session: A persistent interaction context with negotiated parameters (budget, privacy) that eliminates the need to re-transmit history

LLM-as-judge: An evaluation methodology where a strong LLM (here, Gemini) scores the outputs of other models based on criteria like correctness and completeness

Ollama: A framework for running large language models locally

JamJet: The specific Rust-based agent runtime where LDP was implemented as a plugin adapter