Title: Capability Advertising as a Market for Lemons: Establishing a Trust Layer for Heterogeneous Agent Networks

Abstract

As Large Language Model (LLM) agents increasingly delegate tasks among themselves, communication frameworks like the Model Context Protocol (MCP) and the Agent2Agent (A2A) protocol have emerged to facilitate this interaction. These systems allow agents to publish their functional capabilities, enabling others to invoke them, with public registries of such agents already beginning to surface. However, current protocols operate under the assumption that advertised capabilities are static and objectively true. In reality, an agent’s competence is probabilistic, fluctuates based on input, and shifts when the underlying model undergoes updates. Furthermore, because agents are language models themselves, they can articulate their abilities with absolute confidence while being factually incorrect. Consequently, a requester observes an agent’s claims rather than its actual potential, lacking a principled method to distinguish between reliable providers and fluent impostors.

We posit that these challenges stem from a single root cause: the "market for lemons." When quality is obscured and claims are inexpensive to make, high-quality and low-quality providers become indistinguishable. This dynamic ensures that honest reliability is not rewarded, causing the market to deteriorate toward its least competent participants. While economics proposes three remedies—signaling, screening, and reputation—none are currently integrated into agent protocols.

This paper presents four key contributions: (1) a failure taxonomy that identifies "confident-wrong" behavior as a non-adversarial, correlated subclass of Byzantine faults, which classical fault-tolerance models fail to accurately represent; (2) a market-for-lemons model demonstrating that protocols relying solely on faith admit only a low-trust equilibrium; (3) the Trust Layer, a lightweight, protocol-agnostic narrow waist situated above MCP and A2A. This layer introduces probabilistic capability descriptors, screening mechanisms, and reputation systems, enabling a separating equilibrium when the cost of maintaining an overclaim exceeds the benefits gained from it; and (4) a reliability-composition bound for delegation chains, supported by an end-to-end placement argument. The proposed design requires no model retraining and degrades gracefully in scenarios where trust anchors are missing or compromised.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC