九月十五日

Box CEO Aaron Levie predicts that agent swarms, computer use, new APIs and MCPs, and background workflows could generate 100 times the information volume people previously requested in a single session. He sees continuous recruiting, customer monitoring, transcript analysis, and code review as examples, while deployment, management, and budgeting remain early. He argues that overly broad data access threatens security, while locking everything down sacrifices productivity. Box Shield now offers document-classification controls for agent access; anomaly detection, alerts, and blocking are further capabilities he says Box is developing.

Box CEO Aaron Levie 预测，agent swarm、computer use、新一代 API 与 MCP，以及后台 workflow，可能带来人们过去单次 session 所请求信息量的 100 倍。他列举了持续招聘、客户动态监测、对话记录分析和代码审查等场景，同时认为部署、管理和预算机制仍处于早期。他强调，数据访问权限过宽会威胁安全，完全锁死又会牺牲生产力。Box Shield 已提供基于文档分类的 agent 访问控制；他表示，Box 还在开发异常检测、告警和阻断能力。

Vercel CEO Guillermo Rauch argues that agents depend on their proof-checkers, compilers, type systems, and linters, pointing to shadcn/lint for enforcing design-system rules and calling verifiers plus skills the new frameworks. He says Google Cloud Run creator Steren is joining Vercel to lead Fluid compute products across Functions, Containers, Sandbox, and Builds, as agents demand new compute primitives. Rauch also reports that fx auto-upgrades, uses ctrl+g to restart and resume a chat, and runs much faster in long sessions with version 0.0.10.

Vercel CEO Guillermo Rauch 认为，agent 的表现取决于配套的证明检查器、编译器、类型系统和 linter。他以 shadcn/lint 执行设计系统规则为例，把验证工具加 skills 视为新的框架。他表示，Google Cloud Run 创造者 Steren 将加入 Vercel，领导涵盖 Functions、Containers、Sandbox 和 Builds 的 Fluid 计算产品线，以应对 agent 对新计算原语的需求。Rauch 还介绍，fx 支持自动升级，通过 ctrl+g 重启并恢复聊天，0.0.10 版本在长 session 中的速度也明显提高。

Anthropic Claude Code builder Boris Cherny says Claude Mods are landing, with a community member already building a Tetris-in-Claude mod. He links an issue containing the community update, technical details, and demos; the post does not specify the broader release scope.

Anthropic Claude Code 开发者 Boris Cherny 表示，Claude Mods 正在陆续上线，社区已经有人制作了 Tetris-in-Claude mod。他链接的 issue 包含社区进展、技术细节和 demo，但帖子没有说明更广泛的发布范围。

Google VP Josh Woodward says Gemini’s power-user group has tested more than 20 features in the two months since it launched. A new cohort is receiving early access to upcoming Daily Brief and Personal Intelligence features, although the post does not describe the changes themselves.

Google VP Josh Woodward 表示，Gemini 深度用户测试组成立两个月以来，已测试超过 20 项功能。新一批用户正在提前体验 Daily Brief 和 Personal Intelligence 的后续功能，但帖子没有披露具体改动。

FirstMark investor Matt Turck argues that AI progress will not slow down. He attributes that view to the number of competing players, strong economic incentives, and prisoner’s-dilemma dynamics both within countries and internationally.

FirstMark 投资人 Matt Turck 认为，AI 进展不会放慢。他的理由是竞争者众多、经济激励强劲，而且各国内部及国际层面都存在囚徒困境式的竞争。

#### How we contain Claude across products

Anthropic frames agent deployment risk as a blast-radius problem: model safeguards reduce the chance of failure, while stronger capabilities and broader access increase the potential damage. Its three Claude products therefore use different isolation patterns: ephemeral gVisor containers for claude.ai, an OS-level human-in-the-loop sandbox for Claude Code, and a local VM for Claude Cowork. The post reports that users approved about 93% of permission prompts, while Claude Code's sandbox reduced those prompts by 84%, evidence that repeated approval dialogs are a weak primary defense. Anthropic's incidents also show why domain allowlists are capability grants rather than simple destination filters: a malicious file used an attacker-controlled API key to upload workspace files through the permitted Anthropic API. The fix validates a VM’s own provisioned session token through an intercepting proxy. The practical rule is concise: "The deterministic boundary is what gets hit when everything probabilistic misses."

Anthropic 将 agent 部署风险定义为 blast radius 问题：模型防护降低故障发生的概率，但能力增强和访问范围扩大，会让潜在损害持续上升。因此，三款 Claude 产品采用不同的隔离方式：claude.ai 使用临时 gVisor container，Claude Code 使用由人参与监督的操作系统级 sandbox，Claude Cowork 则使用本地 VM。文章称，用户大约会批准 93% 的权限提示，而 Claude Code 的 sandbox 将此类提示减少了 84%，说明反复弹出的批准窗口并不适合作为主要防线。Anthropic 披露的事故也表明，域名 allowlist 不是简单的目的地过滤器，而是一种能力授权：恶意文件曾利用攻击者控制的 API key，通过被允许的 Anthropic API 上传 workspace 文件。修复方案是在拦截代理中验证为 VM 配置的专属 session token。最实用的原则是：“当所有概率性防护都失效时，真正承受冲击的是确定性的边界。”

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

Anthropic Engineering’s April postmortem attributes Claude Code quality complaints to three independent product-layer changes, not the API or inference layer, and says all were resolved by April 20 in v2.1.116. First, changing default reasoning effort from high to medium reduced latency but also intelligence, a tradeoff Anthropic calls "the wrong tradeoff"; the restored defaults described in the post are xhigh for Opus 4.7 and high for other models. Second, a stale-session caching bug repeatedly deleted older reasoning after an hour of inactivity, causing forgetfulness, repetition, cache misses, and faster usage-limit depletion. Third, a system instruction limiting text between tool calls to 25 words and final answers to 100 words produced a 3% drop on a broader evaluation and was reverted. Anthropic plans public-build dogfooding, per-model evaluation and ablation for prompt changes, soak periods, gradual rollouts, and stronger code review with cross-repository context.

Anthropic Engineering 的 4 月复盘指出，Claude Code 的质量投诉来自三个彼此独立的产品层改动，并非 API 或推理层问题；截至 4 月 20 日的 v2.1.116，三项问题都已修复。第一，将默认 reasoning effort 从 high 调到 medium 虽然降低了延迟，却也降低了智能水平，Anthropic 称这是“错误的取舍”；文章所述的恢复后配置是 Opus 4.7 默认使用 xhigh，其他模型默认使用 high。第二，一个针对闲置 session 的缓存 bug，会在闲置超过一小时后反复删除旧 reasoning，导致遗忘、重复、cache miss 和 usage limit 更快耗尽。第三，一条限制工具调用之间文本不超过 25 个词、最终回答不超过 100 个词的 system prompt，在更广泛的评测中造成 3% 的下降，随后被撤回。Anthropic 计划扩大公开版本的内部 dogfooding，对每次 prompt 改动执行逐模型评测和 ablation，并加入 soak period、渐进式发布，以及带跨 repository context 的更强代码审查。

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

Anthropic's Managed Agents service decomposes a long-running agent into a durable append-only session log, a replaceable harness that calls Claude and routes tools, and disposable sandboxes that execute code and edit files. This removes the original single-container "pet": a failed sandbox can be reprovisioned, a failed harness can reload the event log and resume, and credentials can remain in a vault or resource-specific integration instead of entering the execution environment. Provisioning a sandbox only when a tool call needs one reduced median time to first token by roughly 60% and p95 by more than 90%. The design also keeps full history outside the model's finite context, hence Anthropic's key distinction: "The session is not Claude's context window." Stable interfaces let future harnesses connect many stateless brains to many execution environments without coupling their lifecycles.

Anthropic 的 Managed Agents 服务把长时间运行的 agent 拆成三个部分：持久化、仅追加的 session 日志；可替换的 harness，负责调用 Claude 和分发工具；以及可随时丢弃并重建的 sandbox，负责执行代码和修改文件。这消除了早期单 container 架构中的“宠物服务器”：sandbox 故障后可以重新创建，harness 故障后可以重载事件日志并继续，凭证则留在 vault 或绑定到特定资源的集成中，不进入执行环境。只有在工具调用确实需要时才启动 sandbox，使 p50 首 token 时间下降约 60%，p95 下降超过 90%。完整历史也被保存在模型有限 context 之外，因此 Anthropic 强调：“session 不是 Claude 的 context window。”借助稳定接口，未来不同的 harness 可以把多个无状态 brain 连接到多个执行环境，而不必绑定彼此的生命周期。

#### Claude Code now supports artifacts

Claude Blog’s June 18 announcement introduces the ability to publish session work as live artifacts such as PR walkthroughs, system maps, dashboards, and release checklists. An artifact draws from the session's codebase, connectors, and conversation, then updates at the same URL with version history and restoration. This gives teams a shared view of an investigation or review as the work progresses, without assembling a separate reporting pipeline. Anthropic says the feature is "private to its author by default"; sharing is limited to authenticated members of the organization, while admins receive org-level controls, retention policies, role scoping, and compliance API visibility. The announcement lists beta availability to Claude Team and Enterprise organizations, with creation from the CLI and desktop app and viewing in any browser.

Claude Blog 6 月 18 日的公告介绍，Claude Code 可以把 session 中的工作发布为实时 artifact，例如 PR walkthrough、系统地图、dashboard 和 release checklist。artifact 会利用 session 里的 codebase、connector 和对话 context，在同一个 URL 上持续更新，并保留版本历史与恢复能力。这样，团队可以随着工作推进共享调查或审查结果，无须另建一套汇报流程。Anthropic 表示，该功能“默认仅作者本人可见”；分享范围限于组织内已认证成员，管理员则拥有组织级开关、保留策略、角色范围控制和 compliance API 可见性。公告说明，该功能面向 Claude Team 与 Enterprise 组织提供 beta，可从 CLI 和桌面应用创建，并在任意浏览器中查看。

通过 Follow Builders skill 生成。