Capture the Flag (CTF) challenges are a well-established benchmark for evaluating automated cybersecurity systems. With the introduction of powerful tools such as Large Language Models (LLMs), interest in solving these challenges automatically is rapidly growing. These models present both architectural strengths and systematic limitations that hinder their practical applicability. This work proposes and evaluates a neuro-symbolic multi-agent architecture for automatically resolving cryptographic CTF challenges. The architecture is organised into three layers: a Coordinator agent that manages the entire flow from challenge analysis to attack hypothesis; a Belief-Desire-Intention (BDI) agent that structures the reasoning, generation, validation, execution, and debugging cycle; and an embodied Explorer agent that perceives and acts within an isolated Docker sandbox.