AI专题·Agent：智能体基建厚积薄发，商业化应用曙光乍现

Investment Rating - The report does not explicitly state an investment rating for the industry Core Insights - The AI development stage is transitioning from reasoning models to intelligent agents, with models increasingly capable of understanding goals and possessing external memory and reasoning abilities [4][6][8] - The AI large model capabilities are advancing through three expansion curves: pre-training, post-training, and test-time, with a shift in focus from training to testing, leading to decreased reliance on large-scale clusters and increased demand for reasoning computing power [4][30][33] - The infrastructure for intelligent agents is rapidly developing, with a flourishing developer ecosystem supported by emerging communication protocols and development tools [4][22][72] - Initial products are accelerating revenue generation, with intelligent agent applications beginning to see commercial viability, showcasing significant annual recurring revenue (ARR) potential [4][24][25] Summary by Sections 1. AI Development Stage - AI is evolving from a reasoning model to an intelligent agent, with five levels of AI capability defined: Chatbot, Reasoner, Agent, Innovator, and Organization [6][7][8] - Current AI products are in a transitional phase, with intelligent agents expected to revolutionize interaction efficiency [8][11] 2. Agent Model Layer - The foundational intelligence level of models is improving, with reasoning capabilities becoming central [4][12][15] - The intelligent agent formula is defined as Agent = LLM + Memory + Planning + Tool use, emphasizing the importance of these components for autonomous decision-making [12][15] 3. Agent Intermediate Layer - The intermediate layer tools are rapidly being constructed, with data governance and model management becoming critical topics [24][25] - The emergence of tools like MCP and A2A is facilitating the development of a robust developer ecosystem [72][74] 4. Agent Application Layer - Initial products such as Cursor and Glean are generating significant revenue, indicating a promising commercial future for intelligent agents [4][24][25] - The report highlights the need for further enhancements in planning capabilities, memory, and multi-modal understanding to unlock monetization potential [4][24] 5. Related Companies - Key players in the industry include Nvidia and Broadcom for reasoning computing power, Google and Snowflake for intermediate tools and data layers, and Salesforce, SAP, and Shopify for downstream applications [4][24]