Trustworthy Multi-Agent Systems: Mitigating Semantic Drift with the Argent Signaling Protocol

Anantha Sharma

Keywords: Multi-Agent Systems, Explainable AI, Information Geometry, Trustworthy AI, Semantic Drift, Large Language Models

TL;DR: Agents communicate through a constrained, provenance-aware latent workspace, while retaining private latent state to preserve diversity and prevent consensus collapse

Abstract: As Large Language Models (LLMs) move from single-turn prompting to orchestrated Multi-Agent Systems (MAS), free-form agent-to-agent communication makes uncertainty, provenance, and failure modes difficult to audit. This is especially problematic in regulated settings that emphasize explainability, traceability, and measurable risk controls. We present the Argent Signaling Protocol (ASP), a lightweight structured header for agent-to-agent payloads that externalizes calibrated confidence, grounding, stochasticity, and indexed assumptions. We pair ASP with an orchestration layer that monitors bounded distributional drift through Jensen-Shannon divergence (JSD) and escalates through explicit circuit-breaking telemetry instead of silent prompt retries. We also outline an Orthogonal Latent Message Bus (OLMB) architecture in which ASP semantics are learned as orthogonal latent channels and exchanged over a shared latent bus through cross-attention taps and device-to-device transport. The contribution of this paper is architectural rather than empirical: we define the protocol, motivate an information-geometric monitoring abstraction, and specify a companion artifact bundle for future evaluation. The resulting design is intended to make semantic drift more visible, auditable, and actionable for human operators supervising enterprise MAS.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 9