Dynamic Cog-Skill: A Dual-Layer Metacognitive Architecture for Collective Intelligence in LLM Multi-Agent Systems

This workshop position paper presents the Cog-Skill dual-layer architecture, addressing fundamental limitations in current AI systems that lack intelligent dynamic switching between rapid response and deep reasoning. Unlike traditional cognitive architectures (ACT-R, SOAR) that rely on functional module division, our approach establishes temporal scale separation as the core design principle, treating fast processing (Skill) and deliberative reasoning (Cog) as activation patterns of the same cognitive network under different temporal situations. The framework includes dynamic cognitive bottleneck switching mechanisms and unified cognitive state control, avoiding the communication overhead of discrete module switching. This position paper presents the theoretical foundations without empirical validation, seeking community feedback on this temporal cognition approach for building more adaptive AI systems.