Devi Parikh 探讨 AI 代理如何将网页交互从手动点击转变为高级任务描述，并分享他从计算机视觉到创立 Yutori 的历程。
Devi Parikh discusses how AI agents will transform web interaction from manual clicking to high-level task description, and shares his journey from computer vision to founding Yutori.
要点 · TL;DR
网页交互从点击转变为向 AI 代理描述任务。 Web interaction shifts from clicking to describing tasks to AI agents.
对于浏览器代理，视觉截图比解析 DOM 更可靠。 Visual screenshots are more reliable than DOM parsing for browser agents.
Scout 监控网页以检测用户定义的事件并通过邮件通知。 Scouts monitor the web for user-defined events and notify via email.
核心观点 · Key points
网络交互将从点击按钮转变为向AI智能体描述任务。 Web interaction will shift from clicking buttons to describing tasks to AI agents.
构建可靠的网络智能体需要从模型到产品的全栈创新。 Building reliable web agents requires innovation across the full stack, from models to products.
使用视觉截图训练浏览器智能体比解析DOM更可靠，泛化性更好。 Using visual screenshots for browser agents is more reliable than parsing DOM for generalization.
对通用视觉语言模型进行后训练用于浏览器使用，可提升可靠性、降低成本并提高能力上限。 Post-training a general VLM for browser use improves reliability, cost, and capability ceiling.
Scouts监控用户定义的事件，通过邮件通知，在后台持续运行。 Scouts monitor the web for user-defined events and notify via email, operating in the background.
反共识 · Contrarian takes
Arc、Comet等AI浏览器增强现有浏览器，但未来是无浏览器的智能体。 AI browsers like Arc and Comet enhance existing browsers, but the future is browserless agents.
网站应根据用户意图不同而呈现不同内容，而非对所有人显示相同页面。 Websites should render differently per user intent, not show the same page to everyone.
智能体应在云端运行而非本地设备，以便在合上笔记本电脑后持续工作。 Agents should run in the cloud, not on-device, to persist when laptops are closed.
多智能体系统对于扩展后台网络任务而不占用设备资源至关重要。 Multi-agent systems are essential for scaling background web tasks without overwhelming the device.
强化学习可以发现监督微调无法学到的网络导航技巧。 Reinforcement learning can discover hacky web navigation tricks that supervised fine-tuning misses.
本期章节 · Chapters（共 26）
AI代理的愿景与介绍Introduction and vision for AI agents
机器人背景与创始团队Robotics background and founding team
赌注与愿景Stakes and vision
空间概览与差异化Space overview and differentiation
产品路径：用例vs平台Product approach: use cases vs platform
Scout产品概览Scout product overview
技术挑战与架构Technical challenges and architecture
网络监控与浏览器使用细节Nuances of web monitoring and browser use
MCP服务器示例MCP servers examples
浏览器使用模型方法Browser Use Model Approach
后训练策略Post-training Strategy
基础模型选择Base Model Choice
系统演进方法System Evolution Approach
评估与用户反馈Evaluation and User Feedback
Scout作为起点，而非报告生成器Scouts as Starting Point, Not Just Report Generator
反馈作为代理循环的一部分Feedback as Part of Agent Loop
浏览器使用的领域特异性vs泛化Domain Specificity vs. Generalization in Browser Use
数据扩展与平台期Data Scaling and Plateauing
结构化RL调优Structuring RL Tuning
超越Scout：未来方向Beyond Scout: Future Directions
个人用例Personal Use Cases
Scout的用例Use Cases for Scout
摄取架构与工具Ingestion Architecture and Tools
多代理协作Multi-Agent Collaboration
跨运行的冗余与学习Redundancy and Learning Across Runs
代理工作流中平衡冗余与动态性Balancing redundancy and dynamism in agentic workflows