从最顶级的30个AI Agent产品里，看懂了这三个趋势

Core Insights - The article discusses the rapid development and adoption of AI agents, highlighting that 62% of companies are experimenting with them, as reported by McKinsey [1][2]. Group 1: Product Development - Most AI agent products are expected to be released between 2024 and 2025, focusing on three main directions: chat-based agents, enterprise automation platforms, and browser/GUI agents [5][7]. - The majority of agents rely on models from leading labs like Anthropic, Google, and OpenAI, with a significant number depending on GPT, Claude, or Gemini series [7][36]. - Out of 30 agents analyzed, 23 are completely closed-source, indicating a trend of "open framework, closed product" in the industry [9][10]. Group 2: Functionality and Autonomy - There is a significant differentiation in the functionality of the 30 agents, particularly in their action space and autonomy levels [11][12]. - Enterprise workflow agents connect to systems like CRM and databases, while command-line interface (CLI) agents operate directly on file systems [13][14]. - User interfaces vary, with enterprise platforms adopting visual orchestration interfaces, while consumer-level agents primarily use chat interfaces [16][17]. - Autonomy levels differ, with many agents functioning as "round assistants" that require user input after each action, while browser agents can operate independently once given a command [18][19]. Group 3: Responsibility and Transparency - The article notes that while the interface layer is becoming standardized, the identity layer is diverging, with many agents not disclosing their AI identity to users [24][25]. - Most agents do not provide clear mechanisms for verifying their identity, raising concerns about user awareness of AI interactions [30][32]. - There is a lack of uniformity in monitoring and oversight mechanisms across different products, leading to varying levels of transparency [33][34]. Group 4: Structural Trends - The report identifies three structural trends: uneven security disclosure, high concentration of foundational models, and a fragmented responsibility chain [34][36][38]. - The reliance on a few foundational models creates efficiency but also poses single-point risks that could affect downstream systems [36][37]. - The distributed architecture of agent systems complicates accountability, as no single entity is responsible for the complete behavior of the agents [38][39].