八月十三日

Claude expanded persistent browser sessions across desktop, web, and mobile, with conversations, skills, and connectors attached to the account instead of a single device. The same rollout also stressed the security tradeoff: browser agents are useful precisely because they can act across tools and pages, but Anthropic warned that hidden instructions inside webpages can still try to steer them.

Claude 把浏览器里的持久会话扩展到了 desktop、web 和 mobile，conversation、skill 与 connector 都跟着账号走，不再绑在单一设备上。同一轮更新也强调了安全取舍：browser agent 之所以有价值，正是因为它能跨工具和页面行动，但 Anthropic 也提醒，网页里隐藏的指令依然可能试图操纵它。

Josh Woodward announced another wave of Gemini integrations across everyday services including Angi, Fever, GetYourGuide, Granola, iHeartRadio, OpenTable, Pandora, Ticketmaster, Wix, Zocdoc, and Zoho. The message is less about one feature than about distribution: Google wants Gemini to become an action layer stitched into the software people already use.

Josh Woodward 宣布 Gemini 又接入了一批日常服务，包括 Angi、Fever、GetYourGuide、Granola、iHeartRadio、OpenTable、Pandora、Ticketmaster、Wix、Zocdoc 和 Zoho。重点不只是单个功能，而是分发策略：Google 想把 Gemini 变成嵌入现有软件栈里的 action layer。

Peter Yang argued that computing is moving from manual keyboard-and-mouse work toward directing cloud agents with voice. His core claim is that the interface shift matters because delegation, not typing speed, becomes the bottleneck once agents can keep working remotely after the human stops.

Peter Yang 认为，计算机使用方式正从手动键盘鼠标操作，转向用语音指挥云端 agent。核心判断是，界面的变化之所以重要，是因为一旦 agent 能在用户离开后继续远程工作，真正的瓶颈就不再是打字速度，而是如何做有效 delegation。

Madhu Guru made the application-layer case plainly: models will keep getting cheaper, better, and more local, so the defensible work shifts to understanding a user's workflow deeply enough to redesign it. He also noted that when more people can build AI products, the premium rises on execution quality rather than mere access to models.

Madhu Guru 很直接地给出了 application layer 的逻辑：模型会持续变得更便宜、更强，也更适合本地化，因此真正可防御的价值会转向对用户 workflow 的深度理解，以及围绕它做重构。他还指出，当越来越多人都能做 AI 产品时，稀缺的就不再是模型访问权，而是执行质量。

Aaron Levie framed the latest DeepSeek and Grok releases as Jevons paradox in action: better models at dramatically lower cost do not shrink demand, they widen the set of tasks companies will attempt. His practical implication is that more model choice increases the value of the application layer that routes work by price, quality, and task shape.

Aaron Levie 把最新的 DeepSeek 与 Grok 发布称作 AI 版的 Jevons paradox：模型能力提升且成本大幅下降，不会压缩需求，反而会扩大企业愿意交给 AI 处理的任务集合。落到实践层面，他的结论是，模型选择越多，按价格、质量与任务形态来做路由的 application layer 就越有价值。

Peter Steinberger reduced the coding-agent timeline to a clean progression: first CLI, then apps, now services, web interfaces, and cloud sessions. It is a useful shorthand for the broader shift from local copilots toward durable agent systems that live online and keep state between sessions.

Peter Steinberger 用一句很清楚的话概括了 coding agent 的演进路径：先是 CLI，然后是 app，现在轮到 service、web interface 和 cloud session。这也对应更大的产品迁移：agent 正从本地副驾驶，变成在线常驻、能跨会话保持状态的系统。

Garry Tan released GBrain v0.45.6.0 with 17 new skills and argued that a personal AI should run as a separate agent with its own memory repository, not inside the main coding agent. The architectural point is that long-lived personal context may be more useful as an independent layer that can work with Codex or Claude Code rather than being trapped inside either one.

Garry Tan 发布了 GBrain v0.45.6.0，新增 17 个 skill，并主张 personal AI 应该作为独立 agent 运行，拥有自己的 memory repository，而不是塞进主 coding agent 里。这个观点背后的架构含义是：长期积累的个人上下文，更适合作为一个独立层存在，与 Codex 或 Claude Code 协作，而不是被困在某一个 agent 内部。

#### An update on recent Claude Code quality reports

Anthropic said the recent Claude Code quality complaints came from three product-layer changes rather than an API or inference regression, and that all three were fixed by April 20 in v2.1.116. The failures were a March 4 default shift from high to medium reasoning effort, a March 26 stale-session bug that repeatedly dropped earlier reasoning, and an April 16 prompt change meant to reduce verbosity that hurt coding quality and caused a 3% drop on one broader eval. Anthropic also said it will tighten prompt review, run broader per-model evals and ablations, and use slower rollouts for changes that could trade off against intelligence.

Anthropic 表示，近期 Claude Code 质量下降的反馈，源头是三项产品层改动，而不是 API 或 inference 层回退；截至 4 月 20 日的 v2.1.116，这三个问题都已修复。具体包括：3 月 4 日把默认 reasoning effort 从 high 调到 medium；3 月 26 日的 stale-session bug 会反复丢弃更早的 reasoning；以及 4 月 16 日为减少冗长输出而加入的 prompt 改动，在一项更广泛的 eval 上带来了 3% 下滑。Anthropic 还表示，后续会收紧 prompt review，扩大逐模型 eval 和 ablation，并对可能损伤 intelligence 的变更采用更慢的 rollout。

#### Scaling Managed Agents: Decoupling the brain from the hands

Anthropic's deeper engineering write-up explains Managed Agents as three stable interfaces: a session log, a harness that calls Claude and routes tools, and replaceable execution environments. The key move is separating the agent's brain from its hands so crashed harnesses can resume from an external session log and credentials never need to live inside the generated-code sandbox. Anthropic says that decoupling dropped p50 time-to-first-token by roughly 60% and p95 by more than 90%, while making it easier to connect agents to customer-controlled environments and multiple execution targets.

Anthropic 在更深入的工程文章里，把 Managed Agents 拆成三个稳定 interface：session log、负责调用 Claude 与路由 tool 的 harness，以及可替换的执行环境。核心变化是把 agent 的 brain 和 hands 分离，这样 harness 崩溃后可以从外部 session log 恢复，credential 也不必再进入运行生成代码的 sandbox。Anthropic 表示，这种解耦让 p50 time-to-first-token 下降约 60%，p95 下降超过 90%，同时也更容易把 agent 接到客户控制的环境和多个 execution target 上。

#### New in Claude Managed Agents: self-hosted sandboxes and MCP tunnels

Claude Managed Agents can now run tool execution inside infrastructure the customer controls, or through managed providers such as Cloudflare, Daytona, Modal, and Vercel. Anthropic kept orchestration and recovery on its side, but moved code execution, files, packages, services, and sensitive data back inside the enterprise perimeter; MCP tunnels add private tool access through a single outbound, end-to-end encrypted connection without public endpoints or inbound firewall rules. Together, the product update makes the managed-agent story much more plausible for companies that want agent capability without surrendering network boundaries.

Claude Managed Agents 现在可以把 tool execution 放进客户自己控制的基础设施里，或交给 Cloudflare、Daytona、Modal、Vercel 等托管 provider。Anthropic 保留 orchestration 与 recovery，但把 code execution、文件、package、service 和敏感数据重新放回企业 perimeter；MCP tunnel 则通过单一 outbound、端到端加密连接提供私有 tool access，不需要 public endpoint 或 inbound firewall rule。合在一起看，这次更新让“既要 agent 能力、又不放弃网络边界”的企业部署路径现实了很多。

The Takeaway: Kevin Scott thinks model reasoning is no longer the main bottleneck. The harder problem now is building the surrounding infrastructure that lets agents remember, ask for permission, call tools, and keep working asynchronously across many systems.

Kevin Scott, Microsoft's CTO, describes an agentic web that needs the same kind of open plumbing the early internet needed. He compares MCP to HTTP and NLWeb to HTML, arguing that the real unlock is not one magical model but a common protocol layer that lets agents interact with websites, APIs, and internal enterprise systems without every company shipping its org chart as product architecture. For Microsoft, that means pushing internal systems to speak standard protocols so different agents can operate across them cleanly.

His security framing is equally practical. Agents need identities and entitlements, not just raw tool access. An agent should be able to say which systems a task requires, ask the user for the needed permissions, and stay governable by administrators. Scott's broader point is that open systems do not have to mean weak security; they need better permissioning and clearer boundaries.

On software craft, Scott rejects the idea that agents erase making. He sees coding agents as one more tool transition in a four-decade career and expects developers to keep strong preferences, mix tools irrationally and rationally, and choose different agents for different jobs. His longer-term prediction is that today's synchronous prompt-response loops will give way to asynchronous delegation: "go sort this out" becomes a normal interface, and the agent returns later with a finished result or a clear handoff.

核心结论：Kevin Scott 认为，模型 reasoning 已经不再是主要瓶颈。更难的问题是把周边基础设施补齐，让 agent 能记忆、申请权限、调用工具，并在多个系统之间异步持续工作。

作为 Microsoft CTO，Kevin Scott 描述的是一个需要开放协议层的 agentic web，作用就像早期互联网需要的公共底层一样。他把 MCP 比作 HTTP，把 NLWeb 比作 HTML，意思是关键不在某个神奇模型，而在一层通用协议，让 agent 能与 website、API 以及企业内部系统交互，而不是让每家公司都把自己的组织结构直接映射成产品架构。对 Microsoft 来说，这意味着要推动内部系统讲统一协议，让不同 agent 可以自然跨系统协作。

他对安全的描述也很务实。Agent 需要 identity 和 entitlement，而不只是裸露的 tool access。一个 agent 应该能说明某个任务需要访问哪些系统，向用户申请所需 permission，并持续接受 administrator 的治理。Scott 更大的观点是，开放系统不等于安全薄弱，真正缺的是更好的 permissioning 和更清晰的边界设计。

谈到软件 craft，Scott 并不接受“agent 会消灭创造感”这种说法。在他四十多年的编程经历里，这只是又一次工具迁移；开发者依然会保留强烈偏好，会理性也会不理性地混用工具，并为不同任务选择不同 agent。他更长期的判断是，今天同步式的 prompt-response 循环会让位给异步 delegation：一句 “go sort this out” 会变成常见界面，而 agent 会在一段时间后带着结果或明确交接再回来。