Hybrid Behaviour Tree–LLM Architectures for Latency-Sensitive Agent Control: A Preliminary Mapping Review, Taxonomy, and Implications for Real-Time Game AI

Agentic AI systems support contextual reasoning, long-horizon planning, and adaptive behaviour, but their use of large language models (LLMs) can introduce substantial latency that limits their suitability for immediate decision-making and highly responsive feedback in latency-sensitive real-time systems such as multiplayer combat games. In contrast, real-time game AI commonly employs low-latency symbolic controllers, including Behaviour Trees, Utility Systems, and Finite State Machines, which provide predictable and computationally efficient execution but offer more limited contextual reasoning and semantic flexibility. Emerging hybrid architectures seek to combine these properties by using LLMs for higher-level reasoning while retaining symbolic controllers for latency-sensitive execution. This review synthesizes current approaches to Behaviour Tree–LLM integration and examines their architectural characteristics for latency-sensitive real-time game AI. As an original contribution, we formalize an end-to-end latency model for decision processes and introduce four parameterised temporal regions for relating decision latency to engineering time budgets. We further develop a multidimensional taxonomy of hybrid BT–LLM architectures covering latency, adaptability, grounding, planning, and autonomy, together with cross-cutting LLM deployment characteristics. Across the reviewed literature, hybrid architectures commonly separate LLM-mediated deliberation from symbolic execution pathways, with LLMs contributing to strategic reasoning, planning, policy selection, contextual interpretation, and adaptation, while Behaviour Trees provide executable control.