MetaCogAgent: Metacognitive Multi-Agent LLM Framework with Self-Aware Task Delegation

A novel multi-agent large language model (LLM) framework named MetaCogAgent has been introduced, drawing inspiration from cognitive science's metacognition theory. Each agent within this framework features a Metacognitive Self-Assessment Unit that assesses alignment with task capabilities prior to execution. The framework presents three key innovations: a self-assessment system that combines articulated uncertainty with historical capability data to gauge confidence for each task; an adaptive delegation protocol that directs low-confidence tasks to more suitable agents through inter-agent evaluations; and a capability boundary learning module that continuously refines each agent's perceived competence limits based on feedback. This approach tackles the issue of overconfidence in task execution, prevalent in existing systems that assign roles without considering agents' self-assessment abilities. The paper can be found on arXiv under ID 2605.17292.