Generative large-language models (LLMs) are evolving from passive tools into adaptive cognitive partners. This paper proposes the Hybrid Agency Framework (HAF), a conceptual model that treats the human-LLM dyad as a single, integrated cybernetic system. Drawing from the traditions of joint cognitive systems and control theory, HAF moves beyond the prevalent "blind prompting" paradigm, where users operate with limited insight into the model's state and bear full liability for its outputs. We formalize the collaborative process as a Partially Observable Markov Decision Process (POMDP), providing a principled, control-theoretic foundation for managing interaction. The framework's central contribution is a model of bidirectional alignment, where both the human's expressed preferences and the LLM's policy co-adapt in real-time within a shared cognitive workspace. This contrasts sharply with current one-way alignment methods (e.g., RLHF) that treat user preferences as static targets for model training. Furthermore, HAF integrates a multi-stakeholder governance matrix explicitly mapped to emerging regulatory standards (EU AI Act, NIST AI RMF, ISO/IEC 42001) to distribute accountability. This paper presents the complete conceptual design, including its theoretical underpinnings, a formal POMDP specification, a multifactor skill-retention protocol to mitigate user de-skilling, and a detailed plan for future empirical validation. By framing human-LLM interaction as a controllable cybernetic system, HAF offers a rigorous roadmap toward safer, more transparent, and genuinely collaborative hybrid intelligence.