Recent work on augmented reality (AR) has explored the use of adaptive agents to overcome attentional issues that negatively impact task performance. However, despite positive technical evaluations, adaptive agents have shown no significant improvements to user task performance in AR. Furthermore, previous works have primarily evaluated such agents using abstract tasks. In this paper, we develop an agent that observes user behaviour and performs appropriate actions to mitigate attentional issues in a realistic sense-making task in AR. We employ mixed methods to evaluate our agent in a between-subject experiment (N=60) to understand the agent's effect on user task performance and behaviour. While we find no significant improvements in task performance, our analysis revealed that users’ preferences and trust in the agent affected their receptiveness of the agent's recommendations. We discuss the pitfalls of autonomous agents and highlight the need to shift from designing better Human–AI interactions to better Human–AI collaborations.