Multiple Automated Finance Integration Agents (MAFIA) With Self-Healing

Arya Sarukkai, Shaohui Sun, Wei Dai

Keywords: LLM, agentic AI, AI in finance, collaborative workflow

Abstract: The integration of agentic artificial intelligence (AI) into financial services presents both transformative opportunities and critical challenges. Agentic systems, autonomous AI agents capable of goal-directed reasoning, adaptation, and collaboration, are increasingly being deployed in high-stakes domains such as lending, compliance, and audit. However, the autonomous and evolving nature of these agents raises substantial concerns about reliability, auditability, adversarial robustness, and regulatory compliance. In this paper, we propose a framework for constructing self-healing, modular agentic systems that interoperate within financial institutions while maintaining correctness and oversight: Multiple Automated Finance Integrated Agents (MAFIA). In addition, we introduce the notion of self-healing, a framework that scores and self-corrects based on a rubric scoring technique tailored to finance. We focus on a representative use case where a lending assistant agent is continuously monitored and audited by a consumer compliance agent. Through baseline experiments involving sensitive prompt evaluation and downstream auditing, we assess the system's alignment with institutional constraints. We further propose an advanced self-learning setup in which agent feedback loops enhance system responses over time, reinforcing accuracy and compliance. Our findings illustrate a path toward trustworthy agentic architectures that combine automation with enforceable safeguards, paving the way for the secure deployment of AI agents in finance.

Submission Number: 13