Thinking Fast and Slow in AI: The Role of Metacognition

The work argues that building more flexible, general AI requires adopting a dual-process architecture inspired by Kahneman's thinking fast and slow: a population of system 1 "fast" agents that react using learned patterns and experience, and system 2 "slow" agents that perform deliberate reasoning and search when problems demand solutions beyond routine responses. Incoming problems are dispatched by a metacognitive controller that consults a world model (domain knowledge) and a self-model (records of past actions and solver skills) to decide which agent or combination of agents should act. The architecture emphasizes multi-agent orchestration rather than monolithic models, with explicit mechanisms to represent solvers' competencies and expected performance.

Metacognition is central: monitoring, confidence estimation, meta-reasoning about computational cost versus expected benefit, and learning from outcomes drive decisions to escalate to slow agents, reallocate resources, or request additional information. Specific components include competence profiles, expected-utility tradeoffs for solver selection, and self-model updates to capture skill and failure modes. The proposed framework aims to improve adaptability to uncertainty and distributional shift, balance speed and accuracy, enable transparent explanations of agent choice, and provide a practical blueprint for integrating experience-based heuristics with deliberative problem solving in AI systems.