The Provenance Paradox in Multi-Agent LLM Routing: Delegation Contracts and Attested Identity in LDP

2026-03-15 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, long

Summary

The paper introduces critical extensions to the LLM Delegate Protocol (LDP) to address the "provenance paradox" in multi-agent LLM systems, where routing based on self-reported quality scores systematically selects the worst delegates. It extends LDP with delegation contracts that define explicit objectives, budgets, and failure policies for delegated tasks. A claimed-vs-attested identity model is added to distinguish self-reported quality from verified metrics (runtime_observed, issuer_attested, externally_benchmarked). Furthermore, typed failure semantics replace unstructured error strings, enabling automated recovery. Controlled experiments with 10 simulated delegates and validation with real Claude models demonstrate that self-claimed routing performs worse than random (simulated: 0.55 vs. 0.68; real models: 8.90 vs. 9.30), while attested routing achieves near-optimal performance (0.95 simulated, 9.80 real). These backward-compatible extensions incur sub-microsecond validation overhead.

Key takeaway

For AI Scientists designing multi-agent LLM systems, relying solely on self-reported quality metrics for task routing is detrimental. You should prioritize implementing protocols that support attested identity, such as the extended LDP, to ensure routing selects capable agents based on verified performance. Incorporating delegation contracts and typed failure semantics will further enhance system reliability and auditability, preventing the "provenance paradox" from degrading overall system performance.

Key insights

Unverified self-claimed quality in multi-agent LLM routing leads to systematically selecting the worst delegates.

Principles

• Trustworthy delegation requires explicit governance.
• Self-claimed quality can actively harm outcomes.
• Attestation protects against strategic inflation.

Method

Extend LLM Delegate Protocol (LDP) with delegation contracts for bounded authority, a claimed-vs-attested identity model for verified quality, and typed failure semantics for automated recovery.

In practice

• Use "issuer_attested" or "externally_benchmarked" claims.
• Implement client-side validation for delegation contracts.
• Utilize typed errors for automated failure recovery.

Topics

• Multi-Agent LLM Systems
• LLM Delegate Protocol
• Provenance Paradox
• Delegation Contracts
• Attested Identity

Code references

• sunilp/ldp-protocol
• google/A2A

Best for: AI Scientist, Research Scientist, CTO, AI Engineer, Machine Learning Engineer, AI Architect

Related on AIssential

• LDP: An Identity-Aware Protocol for Multi-Agent LLM Systems — AI-native protocols exposing model properties enhance multi-agent LLM system efficiency, governance, and reliability.
• Further Notes on Our Recent Research on AI Delegation and Long-Horizon Reliability — LLMs can accumulate semantic degradation in long, delegated workflows, despite strong short-horizon performance.
• Trustworthy Agent Network: Trust in Agent Networks Must Be Baked In, Not Bolted On — Trust in LLM-based agent networks must be architected into the system's core design, not retrofitted.
• Verify Before You Commit: Towards Faithful Reasoning in LLM Agents via Self-Auditing — SAVeR improves LLM agent faithfulness by verifying internal beliefs before action, preventing propagation of unsound reasoning.
• Description-Code Inconsistency in Real-world MCP Servers: Measurement, Detection, and Security Implications — Semantic mismatches between LLM tool descriptions and code, termed DCI, pose critical reliability and security threats.
• Adversarial Consensus Protocol (ACP): Validation by Contradiction in Multi-LLM Systems — Real architectural divergence between LLMs, not staged conflict, is key to robust validation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.