Synthetic Consilium: Multi-Agent AI Architecture Against Cognitive Bias Amplication in Executive Decision-Making

Read full text Summarise with VIRAT
Authors: Igor I. Ivitskiy, Ivan Zymbytskiy
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Executive and entrepreneurial decision-making operates under conditions of extreme uncertainty, where cognitive biases produce systematic, measurable distortions in strategic judgment. Single-agent Large Language Models (LLMs), initially positioned as analytical counterweights, demonstrably fail at this task: sycophancy bias, automation bias, and epistemic monoculture cause single-agent systems to inherit, amplify, and formalize human cognitive errors rather than correct them. This paper proposes the Synthetic Consilium, a structured multi-agent AI architecture that enforces adversarial deliberation, cognitive diversity, and dialectical synthesis to neutralize executive cognitive biases. We develop the information-theoretic foundations of this architecture, formalizing multi-agent debiasing through the Diversity Prediction Theorem (a mathematical identity guaranteeing that ensemble error is strictly less than average individual error), Ashby's Law of Requisite Variety, and the epistemic MIMO analogy from telecommunications theory. We demonstrate that multi-agent AI systems are structurally immune to aective conict, the primary mechanism that degrades constructive disagreement in human teams, and that continuous Bayesian calibration of agent roles enables compounding improvements in decision quality across optimization cycles. Empirical evidence from medical multi-agent diagnosis, multi-agent debate benchmarks, and formal verication systems converges on an optimal architecture of three to ve heterogeneous agents with two to three deliberation rounds. The framework connects directly to the Analyze andTweak stages of the M.A.T.H. Framework [1] and extends the signal integrity principles established in [2] from advertising data pipelines to executive cognitive pipelines.

Fields of Study & Subjects

Artificial Intelligence Applications Psychiatry Mental Health Neuroscience Multi-Agent Systems and Negotiation

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN
Document type
Preprint