AgentSquare - Tsinghua University launches a new framework for modular intelligent agent system design and search.

What is AgentSquare?

AgentSquare, developed by a team from Tsinghua University, automatically searches for large language model agents within a modular design space. Based on standardized module interface abstractions, it enables high-speed self-evolution and adaptive evolution of AI agents. The framework comprises four core modules: task planning, commonsense reasoning, tool usage, and memory learning, supporting optimized agent design for different task scenarios. AgentSquare solves the search problem in agent design based on module reorganization, module evolution, and agent evaluation models, significantly improving agent performance and effectively controlling inference costs.

AgentSquare's main functions

• Modular design spaceAgentSquare 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 reorganizationBased on the optimization of the top-level architecture of the intelligent agent, AgentSquare can reorganize existing high-performance modules and explore better intelligent agent designs.
• Module EvolutionAgentSquare explores and generates new module designs at the code level, introducing innovative designs and expanding the design space.
• Performance predictionIntroducing a Surrogate Model to predict agent performance reduces expensive real-time evaluation costs and accelerates the search process.
• Automated searchAgentSquare automatically discovers and optimizes LLM agent designs without human intervention, enabling automated agent design search.

AgentSquare's technical principles

• Modular Intelligent Agent Search (MoLAS)AgentSquare addresses the MoLAS problem and uses a modular approach to automatically optimize LLM agent design.
• Module reorganization mechanismUsing a large language model (LLM) as the reorganization proposer, and based on performance evaluation experience, a new module combination scheme is proposed.
• Module evolution mechanismBased on LLM as a module programmer, and combined with evolutionary meta-prompt, we explore new module designs.
• Performance prediction modelThe performance of the proposed LLM agent is predicted using an in-context surrogate model, reducing evaluation costs.
• Iterative search algorithmAgentSquare uses an iterative search algorithm, combined with module reorganization and module evolution, to discover agent designs with better performance.
• Standardized InterfaceBased on standardized input/output interfaces for different modules, AgentSquare supports seamless integration and replacement between modules, facilitating the rapid integration and testing of new modules.

AgentSquare's 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 serviceIn the field of customer service, AgentSquare designs intelligent agents to automatically answer customer inquiries, process orders, and provide personalized services.
• Intelligent personal assistantAs a personal assistant, AgentSquare helps users manage schedules, remind them of important events, search for information, and perform daily tasks.
• Education and LearningIn the education sector, AgentSquare creates intelligent teaching aids that provide personalized learning plans and tutoring.
• Medical consultationAgentSquare can assist the healthcare industry by using intelligent agents to provide preliminary diagnosis, health consultation, and medical information retrieval.
• Financial Decision SupportIn the financial sector, AgentSquare designs intelligent agents to analyze market trends, provide investment advice, and conduct risk assessments.