Real-time financial fraud detection systems face significant challenges from adversaries' continually evolving attack strategies. Traditional static classifiers fail to adapt to these changes and often lack interpretability, leading to false positives and missed anomalies. This paper proposes a novel framework combining Multi-Agent Reinforcement Learning (MARL) with Explainable Causal Inference for transaction anomaly detection and resolution. A defender agent learns to identify and intercept fraud in an adversarial environment where an attacker agent simulates fraudulent behaviors. The agents interact within a stochastic game setting and are trained using a centralized critic and decentralized policies. A causal inference module constructs a directed acyclic graph over transaction features to enhance interpretability and applies do-calculus and counterfactual reasoning to explain flagged transactions. We implement a scalable, real-time deployment architecture and evaluate the system using simulated and real transaction data. Results demonstrate that our MARL-based agent outperforms static classifiers in adaptability and recall, while the causal module reduces false positives and provides transparent justifications for fraud decisions. This combination of adaptability and explainability makes the system highly suitable for practical deployment in financial institutions..