Interaction with LLM-Based Systems: A Structured Review and Taxonomy of Mechanisms and Autonomy

Large language models (LLMs) are increasingly integrated into interactive systems across domains such as software development, robotics, and education. As these systems evolve from simple chat interfaces to autonomous, tool-using agents, the design of human–LLM interaction becomes critical. This paper presents a structured review of interaction with LLM-based systems, focusing on how prompting and interaction design mediate system behaviour and autonomy in practice. We analysed 87 studies from 2021–2025, identifying key interaction mechanisms and application-specific challenges. Based on this synthesis, we propose a two-dimensional taxonomy that classifies systems by interaction mechanism (conversational exploration, task-oriented assistance, tool-mediated interaction, and agentic workflows) and level of autonomy (advisory systems, guided execution, delegated execution, and high-autonomy execution). The taxonomy is supported by decision rules, worked examples, and a human-centred lens, emphasizing user control, transparency, error handling, and learning. Our review highlights a shift from single-turn prompting to structured multi-step workflows and the need for evaluation that considers both process and outcomes, particularly in safety-critical settings. This work provides a framework for analysing, comparing, and informing the design of human-centred interaction with LLM-based systems.