Modern access control systems must go beyond identity verification to provide accuracy, adaptability, and immediate threat detection. To address these needs, we present Guardian Agent, a multi-layered security architecture combining fingerprint recognition, RFID badge validation, and contextual reasoning powered by large language models (LLMs). The agent conducts a staged verification process where biometric credentials are cross-checked with secure databases while behavioral and environmental indicators are simultaneously analyzed. Unlike static rule-based approaches, the LLM enables dynamic risk assessment by interpreting schedules, usage patterns, and coherence between credentials, thereby enhancing anomaly detection. Experimental evaluation shows promising performance, achieving 99.1% fingerprint matching accuracy and a 96.1% fraud detection rate, while maintaining low latency for real-time decision making.