What is AgentSquare
AgentSquare is an automatic search for large language model agents in a modular design space, launched by a team from Tsinghua University. Based on standardized module interface abstraction, it enables high-speed self-evolution and adaptive evolution of AI agents. The framework includes four core modules: task planning, commonsense reasoning, tool use, and memory learning, supporting optimized design of agents for different task scenarios. AgentSquare is based on module recombination, module evolution, and agent evaluation models to solve the search problem of agent design, significantly improving agent performance and effectively controlling inference costs.
Main features of AgentSquare
- Modular design space: AgentSquare proposes a modular design space containing four basic modules: planning, reasoning, tool use, and memory, enabling researchers to easily build and optimize LLM agents.
- Module recombination: Based on optimization of the agent's top-level architecture, AgentSquare can recombine existing high-performance modules to explore better agent designs.
- Module evolution: AgentSquare explores and generates new module designs at the code level, introducing innovative designs and expanding the design space.
- Performance prediction: Introduces a surrogate model to predict agent performance, reducing expensive real-time evaluation costs and accelerating the search process.
- Automated search: AgentSquare automatically discovers and optimizes LLM agent designs without manual intervention, enabling automated agent design search.
Technical principles of AgentSquare
- Modular LLM Agent Search (MoLAS): AgentSquare is based on the MoLAS problem and automatically optimizes LLM agent design based on a modular approach.
- Module recombination mechanism: Uses a large language model (LLM) as a recombination proposer to propose new module combination schemes based on performance evaluation experience.
- Module evolution mechanism: Uses an LLM as a module programmer, combined with evolutionary meta-prompt, to explore new module designs.
- Performance prediction model: Uses an in-context surrogate model to predict the performance of newly proposed LLM agents, reducing evaluation costs.
- Iterative search algorithm: AgentSquare is based on an iterative search algorithm, combining module recombination and module evolution to discover agent designs with better performance.
- Standardized interfaces: Based on standardizing the input and output interfaces of different modules, AgentSquare supports seamless integration and replacement between modules, facilitating rapid integration and testing of new modules.
AgentSquare project address:
- Project official website：tsinghua-fib-lab.github.io/AgentSquare_website
- GitHub repository：https://github.com/tsinghua-fib-lab/AgentSquare
- arXiv technical paper：https://arxiv.org/pdf/2410.06153
Application scenarios of AgentSquare
- Automated customer service: In the customer service field, AgentSquare designs agents to automatically answer customer inquiries, process orders, and provide personalized services.
- Intelligent personal assistant: As a personal assistant, AgentSquare helps users manage schedules, remind them of important events, search for information, and perform daily tasks.
- Education and learning: In the education field, AgentSquare creates intelligent teaching assistance tools, providing personalized learning plans and tutoring.
- Medical consultation: AgentSquare can assist the healthcare industry by using agents to provide preliminary diagnosis, health consultation, and medical information retrieval.
- Financial Decision Support: In the financial sector, AgentSquare designs agents to analyze market trends, provide investment advice, and assess risk.