The increasing use of artificial intelligence (AI) and reinforcement learning (RL) in cybersecurity introduces the possibility of autonomous cyber agents capable of adapting their behaviour, learning from defensive responses, and conducting attacks with limited human intervention. Traditional defensive mechanisms are often designed around relatively static assumptions and may struggle against adversaries capable of continuous learning and adaptation. This paper proposes CHARLES (Continuous Homeostatic Adversarial Resilience through Learning and Epistemic Sifting), a conceptual defensive architecture designed to improve resilience against autonomous learning-based cyber attackers and AI agents. CHARLES integrates four principal components: an Epistemic Sifting Engine (ESE) for evaluating the reliability and relevance of observed information; an Adaptive State Transformation Engine (ASTE) for dynamically modifying defensive and observable system states; an Adversarial Learning Estimator (ALE) for estimating attacker adaptation and learning behaviour; and an Adversarial Utility Management Engine (AUME) for influencing attacker utility through coordinated defensive actions. The framework is formulated around a continuous defensive loop of observe, estimate, sift, transform, evaluate, adapt, and repeat. A mathematical model is introduced to represent attacker utility, environmental uncertainty, and adaptive defensive objectives. The proposed methodology compares static defence, moving target defence, cyber deception, and the integrated CHARLES architecture using attack success, policy adaptation, uncertainty, latency, resource consumption, and disruption as evaluation metrics. The framework is intended as a testable research architecture for simulation-based investigation of adaptive cyber defence against autonomous adversarial agents.