Proceedings of the Annual Meeting of the Cognitive Science Society

Large language model (LLM) agents have shown strong capabilities in long-horizon tasks such as cyber threat analysis. However, despite competent step-level reasoning, existing agents often fail to maintain coherent behavior over extended investigation trajectories, leading to accumulated process-level errors. Prior approaches mainly improve individual reasoning steps through prompting or reflection, but provide limited support for sustained goal maintenance, strategy regulation, and execution monitoring. We propose Lumos, a cognitively inspired agent framework that realizes an explicit executive-control loop via process-level control mechanisms, including goal-centric memory for long-term goal maintenance, failure-driven strategy adaptation for inhibitory regulation, and action and output control for explicit monitoring and verification. Experiments on real-world security investigation scenarios demonstrate that Lumos consistently outperforms strong baselines, improving long-horizon stability and overall task success.