Group 1 - The development of AI Agents has accelerated significantly since 2025, with notable acquisitions and funding rounds, such as OpenAI's $3 billion acquisition of Windsurf and Anysphere's $900 million funding round, valuing Cursor at $9 billion [1][3] - The emergence of various platforms and tools, such as MindOS and Second Me, indicates a growing trend towards creating personalized AI Agents, reflecting a shift in the industry towards more accessible development [4][6] - The definition of AI Agents has evolved, now characterized by their ability to perform tasks independently, driven by large language models, and equipped with memory systems and user interaction interfaces [6][8] Group 2 - The integration of reasoning models and Reinforcement Fine-Tuning (RFT) technology has enabled AI Agents to learn and adapt in specific domains, marking a significant advancement in their capabilities [8][15] - The distinction between traditional reinforcement learning Agents and modern AI Agents lies in their ability to learn from environments, with the latter now capable of autonomous learning and exploration [12][14] - The competitive landscape for AI Agents is shifting, with companies like Cursor and Windsurf leading the charge due to their deeper understanding of environments and user needs [18][20] Group 3 - The rise of AI Agents has created both opportunities and challenges for entrepreneurs, as the market becomes saturated with service-oriented Agents, making it difficult for new entrants to find unique value propositions [22][23] - The importance of model capabilities, engineering skills, and data barriers is highlighted as key competitive advantages in the AI Agent space, with the performance of models like Claude Sonnet 3.7 being pivotal for success [25][28] - The future of AI Agents may see a convergence of programming tools and general-purpose Agents, as companies like Cursor and Windsurf begin to integrate broader functionalities [31][55] Group 4 - The industry is experiencing a rapid pace of development, with a shift towards faster execution and less emphasis on detailed planning documents, reflecting a more agile approach to product development [64][66] - Despite the excitement around AI Agents, significant challenges remain in achieving widespread adoption and understanding user needs effectively, indicating that the journey towards mainstream usage is still ongoing [68][71] - The MCP protocol, which governs how AI Agents access external information, is still in its early stages and requires industry-wide acceptance to fully realize its potential [71][73]

与市面上多数"通用Agent"不同，OpenAI 的 Deep Research 从诞生那一刻起就被锁定在一件事上—— 通过强化 学习，将搜索、浏览、筛选与整合信息的能力内化为模型的原生技能，直接训练进参数里，而不是仅靠 Prompt工程和外部工程组合 。 那么，OpenAI 是如何把这套复杂技能训练进参数里的？他们在数据筹备、强化微调、安全与记忆管理上又摸 索出了哪些最佳实践？ OpenAI Deep Research团队核心成员Isa Fulford最近在一个访谈中做了分享： 我们认为这个访谈提供了一个透视 OpenAI 构建旗舰智能体 Deep Research 的独特视角，并提供了一些开发实 践经验，因此锦秋基金（ 微信公号锦秋集ID：jqcapital）对本文进行了编译。 01 Deep Research 的起源与目标 OpenAI 团队在强化学习算法刚刚显露锋芒时，放弃了订汉堡、订花那条看似容易衡量的交易型赛道， 转而攻克浏览与知识整合——他们认为整合知识是AGI 必不可少的前置技能， 也因为"纯读取"比"直接 下单"更安全。 数据的质量比数量更重要。 Deep Research 倾向"小而准"： ...