Core Insights - Skillful Craftsman Education Technology (EDTK) has launched an AI communications app named "Sesame Chat" in mainland China, which utilizes a Multi-Agent System (MAS) for adaptive communication skills learning [1][2] - The MAS consists of multiple AI agents that dynamically activate based on user scenarios, enhancing communication effectiveness across various contexts such as dating, professional interactions, and social scenarios [2][8] Company Overview - Skillful Craftsman focuses on educational technology innovation, artificial intelligence skills training, and the digital transformation of educational institutions [1][8] - The company aims to improve teaching effectiveness, student outcomes, and community connections through intelligent learning platforms and teaching management systems [8] Technology and Functionality - The MAS includes five specialized AI agents: - Style Profiler: Identifies communication preferences [3] - Context Interpreter: Analyzes relationship history for appropriate intimacy levels [4] - Knowledge Specialist: Retrieves relevant cultural frameworks [5] - Response Architect: Generates tailored communication options [6] - Quality Validator: Scores responses and presents the best options to users [7] - The AI agents collaborate seamlessly, utilizing dedicated neural networks and deep reinforcement learning to provide customized communication recommendations [8]

Core Insights - The article discusses the construction of an effective multi-agent research system by the Claude team, focusing on system architecture, prompt engineering, and evaluation methods [1][5][12]. Group 1: System Architecture - The Claude team employs a coordinator-worker architecture to manage task allocation and collaboration among multiple agents [5]. - The system utilizes multi-step search instead of static retrieval, dynamically seeking relevant information and adapting to new findings [8]. - The main agent decomposes queries and initiates specialized subagents, each with its own tools, prompts, and memory, integrating their results [13]. Group 2: Performance Metrics - The multi-agent system significantly enhances performance in research tasks, achieving over 90% success in internal evaluations compared to single-agent models [14]. - The latest Claude model has doubled token efficiency compared to previous versions, with token costs being 15 times higher than standard chat [15]. Group 3: Task Management and Optimization - The team uses heuristic methods for prompt design to optimize agent behavior, focusing on task complexity, clarity of delegation, tool selection, and thinking strategies [16]. - The main agent assigns tasks by breaking down queries into sub-tasks with clear goals and expected outputs, adjusting the scale of work based on task complexity [17]. Group 4: Evaluation Methods - The team employs small sample evaluations to test agent performance early in development, significantly improving success rates [21]. - A large language model (LLM) is used as a judge to assess outputs based on criteria such as factual accuracy and source quality [22][23]. - Human evaluators play a crucial role in identifying anomalies that automated scoring may miss, ensuring the reliability of the system [24]. Group 5: Challenges and Recommendations - The article highlights the "butterfly effect" in agent systems, where minor changes can lead to significant behavioral shifts, necessitating robust recovery systems [29]. - The team introduces asynchronous execution to enhance parallel processing, although it presents challenges in result coordination and error propagation [30]. - Recommendations include focusing on end-state evaluations rather than step-by-step analysis and managing long-term dialogue effectively [31].