# Agents Thinking Fast and Slow: A Talker-Reasoner Architecture

**Authors:** Konstantina Christakopoulou, Shibl Mourad, Maja Matarić  
**arXiv:** [2410.08328](https://arxiv.org/abs/2410.08328)  
**Date:** 2024-10-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2410.08328v1/

## Key points

- Large language models (LLMs) have transformed agent-user interactions through natural conversation
- Agents now balance conversing with users and engaging in multi-step reasoning and planning to achieve goals
- Dual-system architecture introduced: "Talker" agent for quick, intuitive responses and "Reasoner" agent for deliberate reasoning, planning, and action execution
- Advantages of Talker-Reasoner architecture include modularity and decreased latency
- Model integrates talking while reasoning/planning and explicit belief modeling for a sophisticated understanding of user behavior
- Incorporates natural language feedback into decision-making process to continuously update beliefs about user goals, plans, motivations, and barriers
- Theoretical underpinnings of the model related to human cognition systems are discussed alongside a practical example - a sleep coaching agent - to demonstrate real-world relevance
- Integration of fast intuitive responses with slower logical reasoning enhances conversational agents' capabilities across various domains

## Summary

In recent years, large language models (LLMs) have revolutionized the way agents interact with users through natural conversation. This has led to a shift in the roles of agents, which now must balance conversing with users and engaging in multi-step reasoning and planning to achieve goals. This dichotomy is akin to the concept of "thinking fast and slow" introduced by Kahneman, where quick, intuitive responses (System 1) are complemented by slower, more logical reasoning and planning (System 2). Building upon this framework, our approach introduces a dual-system architecture consisting of a "Talker" agent responsible for synthesizing conversational responses quickly and intuitively, and a "Reasoner" agent tasked with more deliberate reasoning, planning, and executing actions to drive the agent towards its goals. This Talker-Reasoner architecture offers advantages such as modularity and decreased latency. Drawing inspiration from related work on LLM-driven agents that focus on text-based interactions as well as embodied agents capable of multimodal interactions, our model integrates both talking while reasoning/planning and explicit belief modeling. By incorporating natural language feedback into the agent's decision-making process and continuously updating its beliefs about user goals, plans, motivations, and barriers, we aim to create a more sophisticated understanding of user behavior. In addition to discussing the theoretical underpinnings of our Talker-Reasoner model in relation to human cognition systems, we also ground our discussion in a practical example - a sleep coaching agent - to demonstrate the real-world relevance of our approach. Through this detailed exploration of our novel agent architecture, we aim to showcase how integrating fast intuitive responses with slower logical reasoning can enhance the capabilities of conversational agents in various domains.