Taxonomy Paper Maps Cognitive Capability Gaps Blocking Long-Horizon Agent Reliability Across Five Dimensions

arXiv 2608.02553·low signalXBlueskyLinkedInCopy link

This survey organizes the literature on what still breaks in generative and agentic AI around five dimensions — persistent state modeling, goal-directed autonomy, self-monitoring and control, environment interaction, and learning and adaptation — reviewing advances and recurring limitations in each. It proposes a conceptual Adaptive Cognitive Intelligence Architecture (ACIA) and surveys cognition-centric evaluation directions. Useful as a structured checklist for diagnosing why an agent degrades over extended time horizons, though it introduces no new empirical results.