AI Agent深度（二）：2025 Agent元年，AI从L2向L3发展

Investment Rating - The report suggests that 2025 is a crucial investment window for the Agent sector, emphasizing the need to closely monitor advancements in foundational models, reinforcement learning, and standardized protocols like MCP [2][4]. Core Insights - The report identifies 2025 as the "Agent Year," marking the evolution of AI from L2 (Reasoner) to L3 (Agent), indicating a shift from "thinking" to "acting" driven by technological maturity, product benchmarks, protocol proliferation, and market demand [2][3][39]. - The significance of Agents lies in their potential for deep automation, serving as a pathway to AGI, and reshaping internet entry points, with competition expected to intensify in the second half of 2025 [2][3][61]. - The competitive landscape is characterized by major tech giants dominating the general Agent ecosystem while vertical opportunities remain for niche players with deep domain knowledge [2][3][61]. Summary by Sections 1. Why 2025 is the Agent Year - AI is transitioning from L2 to L3, with key drivers including technological maturity, product validation by industry leaders, and market demand for complex task automation [3][39]. - The definition of an Agent requires four essential components, with the ability to call tools being the most critical differentiator [43][44]. 2. Importance of Agents - Agents enable deep automation, freeing humans from repetitive tasks and allowing focus on higher-value creative work [2][3][49]. - They are pivotal in the journey towards AGI and embodied intelligence, with the potential to redefine how users access information and complete tasks [2][3][61]. 3. Competitive Landscape - The competition in the Agent space is dominated by large tech platforms leveraging their model, data, and ecosystem advantages [2][3]. - Vertical opportunities exist for specialized Agents that integrate deep domain knowledge, although they face long-term threats from general Agents [2][3]. 4. Investment Recommendations - The report advises focusing on the Agent investment window in 2025, tracking advancements in foundational models, reinforcement learning, and the reliability of tool invocation [2][4]. - Long-term investments should be directed towards platform giants with robust foundational models and ecosystems, as they are likely to lead the development of general Agents [2][4]. - Attention should also be given to vertical leaders that have established domain knowledge and clear business models before the full maturity of general Agent capabilities [2][4].