01 / Agent 安全
Claude Code Auto Mode 在补上 agent 执行的安全层
Anthropic Engineering 介绍 Claude Code Auto Mode，用模型分类器替代部分人工权限确认，目标是在少打扰用户的同时拦住越权、误删、外传数据等高风险动作
这说明 coding agent 的竞争点不只是谁更会写代码，还包括谁能在真实电脑环境里安全执行任务，Prompt Injection 也被放进了输入层防护里
来源：Anthropic Engineering：《Claude Code auto mode: a safer way to skip permissions》 https://www.anthropic.com/engineering/claude-code-auto-mode
02 / 真实世界 AI
Waymo 把世界模型带进自动驾驶的长期工程
Waymo 的 Dmitri Dolgov 提到 Waymo Foundation Model 支撑 driver、simulator、critic 三个部分，本质是在真实道路里使用 multimodal world action language model
自动驾驶再次提醒我们，AI 落地不是只发一个模型，而是模型、传感器、仿真、安全验证、城市运营一起长期迭代
来源：Training Data 播客：《Waymo's Dmitri Dolgov: 20 Million Rides and the Road to Full Autonomy》 https://www.youtube.com/playlist?list=PLOhHNjZItNnMm5tdW61JpnyxeYH5NDDx8
03 / 产品趋势
Personal Agents 正从代码场景扩展到知识工作
Peter Yang 的判断很直接，Coding 是第一前沿，Knowledge Work 是第二前沿，Personal Agents 是第三前沿
这条线索值得盯住，AI 入口可能从聊天框变成长期跟随的个人 agent，帮用户跨应用处理资料、日程、写作、沟通和执行
来源：Peter Yang 公开 X 动态：Coding、Knowledge Work、Personal Agents https://x.com/petergyang/status/2051508988936937764
04 / 企业落地
企业 AI agents 需要流程、上下文和变革管理
Box CEO Aaron Levie 认为 OpenAI 和 Anthropic 都在推动企业部署 AI agents，这会很快变成大趋势
但企业落地不是把模型接进去就结束，还要升级 IT 系统、给 agent 上下文、改造 workflow，并重新定义 human-agent relationship
来源：Aaron Levie 公开 X 动态：企业 AI agents 落地趋势 https://x.com/levie/status/2051344780328858040
05 / 开源工具
Vercel 开源 deepsec，把安全审查交给并行 coding agents
Guillermo Rauch 介绍 npx deepsec，一个用于 deep security reviews 的 open-source agent orchestrator，目标是让大量 agents 并行检查代码库
这代表 agent 不只是写代码，也开始承担测试、安全、审计等工程流程，软件团队的交付链路会继续被 AI 重塑
来源：Guillermo Rauch 公开 X 动态：Vercel deepsec https://x.com/rauchg/status/2051386798899888539
英文速记
Auto Mode自动执行模式，让 agent 在安全约束下减少反复请求确认
Prompt Injection提示注入，外部内容试图劫持 AI 的原本任务
World Model世界模型，让 AI 理解环境、物理变化和行为后果
Personal Agents个人智能代理，长期跟随用户处理跨应用任务
Workflow工作流，一套从输入到执行再到确认的业务流程
来源：本期 AI Daily 内容中出现的行业英文表达整理
01 / Agent Safety
Claude Code Auto Mode is about safer hands-off execution
Anthropic’s Auto Mode points to a practical shift: agents should ask for fewer confirmations, but still stop before risky actions
The bar is moving beyond “can it write code?” to “can it work on my machine without breaking things?”
Source: Anthropic Engineering
02 / Real-world AI
Waymo shows what long-horizon AI engineering looks like
Waymo’s foundation model sits across driving, simulation, and review, which is a useful reminder that real-world AI is a system, not a single model
Shipping the model is only one layer. Operations, validation, and feedback loops are where the hard work compounds
Source: Training Data podcast
03 / Product Trend
Personal Agents are starting to move beyond coding
A clean way to read the trend: coding came first, knowledge work is next, and personal agents are the bigger interface shift
The next interface may not be another chat box, but a persistent agent that follows work across apps
Source: Peter Yang on X
04 / Enterprise Adoption
Enterprise agents will win by fitting into real workflows
The enterprise story is moving past demos. The useful agents are the ones that understand process, permissions, and context
The hard part is deciding where the agent acts alone, where it asks, and where a human must stay in control
Source: Aaron Levie on X
05 / Open-source Tools
Vercel deepsec hints at agents as code reviewers, not just code writers
Deepsec is interesting because it puts agents into review work: testing, security, and audit, not just generation
Software teams may soon treat parallel agent review as a normal part of the delivery pipeline
Source: Guillermo Rauch on X
Key Terms
Key terms to keep
Auto ModeA mode where the agent can keep working with fewer prompts, while safety checks stay on
Prompt InjectionHidden instructions in content that try to steer the AI away from the user’s real goal
World ModelA model that predicts how the world or environment changes after actions
Personal AgentsAssistants that stay with a user across repeated tasks and apps
Source: English terms collected from this issue