九月二十四日

Anthropic’s Claude account announced that Claude Marketplace now lets users discover tools, agents, and expert partners in one place. Its examples span connectors and plugins such as Slack and Notion, agents and products from Cursor and CrowdStrike, and service partners including Accenture and Deloitte. The scope reaches beyond adding a tool to a conversation: the announcement puts software discovery and implementation help alongside each other, while inviting builders to list their tools, agents, and services. [Announcement](https://x.com/claudeai/status/2102840851538080172) · [Builder listing invitation](https://x.com/claudeai/status/2102840855258452037) · [Browse the marketplace](https://t.co/MIKvTzg59L) · [List a tool or service](https://t.co/gjX3Xctic2)

Anthropic 的 Claude 官方账号宣布，用户现在可以在 Claude Marketplace 中发现工具、agent 和专业合作伙伴。列举的类别包括 Slack、Notion 等连接器和插件，Cursor、CrowdStrike 等公司的 agent 与产品，以及 Accenture、Deloitte 等服务伙伴。这一入口覆盖的不只是给对话添加工具，也将软件发现与实施服务放在了一起，同时邀请开发者上架自己的工具、agent 和服务。 [发布说明](https://x.com/claudeai/status/2102840851538080172) · [开发者上架邀请](https://x.com/claudeai/status/2102840855258452037) · [浏览市场](https://t.co/MIKvTzg59L) · [提交工具或服务](https://t.co/gjX3Xctic2)

Vercel CEO Guillermo Rauch introduced Drives as an attachable storage companion to Sandbox. He separates an agent into model and orchestration logic, execution tools, and files such as memories, skills, and repositories; his concrete example is a nightly memory-consolidation job that can read and write files without booting the agent’s full computer. In his proposed architecture, Fluid compute runs the orchestration logic, Workflow makes its event log durable, and browsers or sandboxes supply execution, with independent storage completing the separation. Rauch argues that this improves cost, security, and auditability, while explicitly describing Drives as early and still expanding in capability. [Source](https://x.com/rauchg/status/2102820148629614685)

Vercel CEO Guillermo Rauch 发布了 Drives，将其定位为可按需挂载、与 Sandbox 配套使用的存储。他把 agent 拆成模型与调度逻辑、执行工具，以及记忆、skills、代码仓库等文件；一个具体场景是每晚整理记忆时，直接读写文件，无需启动 agent 的整台计算机。在他提出的架构中，Fluid compute 运行调度逻辑，Workflow 持久保存事件日志，浏览器或沙箱负责执行，独立存储则补齐了最后一块。Rauch 认为这种拆分有助于降低成本、改善安全性与可审计性，同时明确表示 Drives 仍处于早期阶段，能力还会继续扩展。 [来源](https://x.com/rauchg/status/2102820148629614685)

OpenAI’s Thibault Sottiaux, whose work spans Codex and ChatGPT, described using voice conversations with ChatGPT to discuss work, check email, write code, and manage his calendar. He says the experience works across the plugin ecosystem, including third-party plugins. This is a concrete account of voice reaching connected work tools, although his post does not specify rollout scope, supported plans, or individual integration requirements. [Source](https://x.com/thsottiaux/status/2102814202117411196)

负责 Codex 和 ChatGPT 相关工作的 OpenAI 成员 Thibault Sottiaux 描述了自己通过语音与 ChatGPT 讨论工作、查看邮件、编写代码和管理日历的使用方式。他表示，这种体验可以调用整个插件生态，包括第三方开发的插件。这提供了一个语音接入工作工具的具体案例，但帖子没有说明开放范围、支持的套餐或各项集成的使用条件。 [来源](https://x.com/thsottiaux/status/2102814202117411196)

AI tutorial and interview creator Peter Yang praised Opus 5.5’s intelligence, speed, writing, conversational feel, and rate limits, then argued that Claude Code still needs strong live voice and browser plus computer use. His distinction is between the model’s capabilities and the surrounding software that lets it act; he believes the best model and best agent tooling currently sit at different labs. These are hands-on judgments, and the posts provide no comparative benchmark; Yang also says computer use is improving. [Model impressions](https://x.com/petergyang/status/2102946952740741394) · [Tooling assessment](https://x.com/petergyang/status/2102952201350254667)

AI 教程与访谈创作者 Peter Yang 肯定了 Opus 5.5 的智能程度、速度、写作能力、对话体验和使用限额，随后指出 Claude Code 仍需要出色的实时语音、浏览器操作和计算机操作能力。他区分的是模型本身的能力，以及让模型真正采取行动的配套软件；在他看来，目前最好的模型与最好的 agent 工具环境分属不同实验室。这些属于使用体验判断，帖子没有给出对比基准测试；Yang 也提到计算机操作能力正在改善。 [模型体验](https://x.com/petergyang/status/2102946952740741394) · [工具环境评价](https://x.com/petergyang/status/2102952201350254667)

Meta Senior Director of AI Madhu Guru argues that people do not have a single preference for how much of a task they want to delegate. Shopping can be enjoyable browsing or an urgent request for the right item, whereas hiring a roofer may mean reviews, phone calls, insurance coordination, inspections, and hidden costs. Responding to Ben Thompson, Guru argues that demand for consumer agents is substantial precisely because some workflows should require almost no effort, while others benefit from more exploration and informed choice. [Source](https://x.com/realmadhuguru/status/2102777931764498536)

Meta AI 高级总监 Madhu Guru 认为，人们愿意把多少工作交给 agent，取决于具体任务，并不存在统一偏好。购物可能是享受闲逛，也可能是急着拿到合适的商品；而找人修屋顶，往往意味着查评价、反复打电话、协调保险、安排检查和识别隐藏费用。在回应 Ben Thompson 时，Guru 指出，消费者对 agent 的需求很大，恰恰是因为有些流程最好几乎不费力，有些流程则需要更多探索空间，帮助人做出知情选择。 [来源](https://x.com/realmadhuguru/status/2102777931764498536)

Ryo Lu, a designer whose background includes Cursor, Notion, and Stripe, challenges the habit of measuring AI’s value through more output, more merged pull requests, and more managed agents. His concern is that unlimited execution can arrive before people develop intention and taste, leaving them busier without making their work more meaningful. He argues for discernment, including knowing what not to build and when to stop, as a central ambition for AI tools. [Source](https://x.com/ryolu_/status/2102933485795369213)

曾参与 Cursor、Notion 和 Stripe 设计工作的 Ryo Lu，质疑把更多产出、更多合并的 PR 和更多受管理的 agent 当作 AI 价值的主要衡量方式。他担心，人们还没有形成清晰的意图与判断力，就先拥有了近乎无限的执行能力，结果只是更忙，工作却未必更有意义。他主张把判断力放在 AI 工具的重要目标中，包括知道什么不值得做，以及什么时候应该停下来。 [来源](https://x.com/ryolu_/status/2102933485795369213)

Anthropic’s Claude Code builder Boris Cherny clarified what Claude’s formal-methods workflow actually does: construct a model of a tricky state machine or race-prone part of a program, find counterexamples, reproduce the suspected bugs, and fix the code. He explicitly distinguishes this targeted process from formally verifying an entire codebase. The useful boundary is that a counterexample identifies a suspected defect; reproduction connects the simplified model to a real failure before the fix. [Source](https://x.com/bcherny/status/2102898067133595992)

Anthropic 的 Claude Code 开发者 Boris Cherny 解释了 Claude 使用形式化方法的具体流程：为复杂状态机或容易出现竞争条件的代码建立模型，寻找反例，复现疑似缺陷，再修复代码。他明确指出，这种针对局部复杂逻辑的做法，并不等于已经对整个代码库完成形式化验证。这里的关键边界是，反例首先指向一个疑似问题，而复现步骤负责确认简化模型中的问题确实对应真实代码的故障。 [来源](https://x.com/bcherny/status/2102898067133595992)

Anthropic Engineering’s “How we contain Claude across products,” written by Max McGuinness, Mikaela Grace, Jiri De Jonghe, Jake Eaton, and Abel Ribbink, argues that limiting an agent’s reachable resources is essential even when model safeguards improve. The article reports that users approved roughly 93% of permission prompts, while Claude Code’s OS-level sandbox reduced prompts by 84%. Its central principle is: “Design for containment at the environment layer first, then steer behavior at the model layer.” A revealing failure involved a malicious workspace file instructing Cowork to upload data to an attacker’s account through the allowed api.anthropic.com domain. Anthropic says it fixed that path with a proxy accepting only the VM’s provisioned session token and blocking headers that enable server-side fetching. The practical lesson is that approving a domain grants access to capabilities behind that domain; isolation also creates an observability tradeoff, because host endpoint monitoring cannot inspect the guest VM and after-the-fact log exports do not provide live monitoring. [Original article](https://www.anthropic.com/engineering/how-we-contain-claude)

Anthropic Engineering 的《How we contain Claude across products》由 Max McGuinness、Mikaela Grace、Jiri De Jonghe、Jake Eaton 和 Abel Ribbink 撰写，强调即使模型防护不断改善，也必须限制 agent 实际能够触及的资源。文章称，用户会批准约 93% 的权限请求，而 Claude Code 的操作系统级沙箱让权限提示减少了 84%。其核心原则是：“先在环境层设计隔离与约束，再在模型层引导行为。”一个有代表性的漏洞是，工作区中的恶意文件诱导 Cowork 经由已获准访问的 api.anthropic.com，将数据上传到攻击者的账号。Anthropic 表示，修复方式是通过代理只接受为该虚拟机配置的会话 token，并阻止能够触发服务端抓取的请求头。实际启示是，批准一个域名，也是在开放该域名背后的能力；隔离还会带来可观测性取舍，因为宿主机的终端监控无法检查虚拟机内部，事后导出日志也不能替代实时监控。 [原文](https://www.anthropic.com/engineering/how-we-contain-claude)

Anthropic Engineering’s “Scaling Managed Agents: Decoupling the brain from the hands,” by Lance Martin, Gabe Cemaj, and Michael Cohen, describes separating the durable session log, the agent’s orchestration loop, and the execution sandbox behind independent interfaces. A failed sandbox becomes a tool error, while a failed orchestration process can restart and recover from the external event log. The design also keeps credentials outside the sandbox running generated code. The key distinction is captured by the article’s heading, “The session is not Claude’s context window”: the log preserves events even when the model’s working context is trimmed or compacted. Anthropic reports roughly 60% lower median time to first token and over 90% lower p95, because inference no longer waits for a sandbox to be provisioned when one is not immediately needed. These are reported results for this architecture, rather than a universal speedup for every agent workload. [Original article](https://www.anthropic.com/engineering/managed-agents)

Anthropic Engineering 的《Scaling Managed Agents: Decoupling the brain from the hands》由 Lance Martin、Gabe Cemaj 和 Michael Cohen 撰写，介绍了如何通过独立接口拆分持久会话日志、agent 调度循环和执行沙箱。沙箱故障会变成一次工具调用错误，而调度进程故障后，可以重启并从外部事件日志恢复。这种设计还把凭据留在运行生成代码的沙箱之外。文章用一个小标题概括了关键区别：“会话不等于 Claude 的上下文窗口。”即使模型的工作上下文被裁剪或压缩，日志仍保留事件。Anthropic 报告称，首个 token 的中位等待时间下降约 60%，p95 下降超过 90%，原因是不需要立即使用沙箱的任务，可以先开始推理，无需等待沙箱就绪。这些数字是该架构下的测量结果，不能直接当作所有 agent 任务的通用提速幅度。 [原文](https://www.anthropic.com/engineering/managed-agents)

Anthropic Engineering’s “An update on recent Claude Code quality reports” is an April postmortem, useful as historical context for the current discussion of agent tooling. It attributes quality problems to three product changes: lowering default reasoning effort, a caching bug that repeatedly removed prior reasoning after a session had been idle for over an hour, and a system-prompt change intended to reduce verbosity. The article says “The API was not impacted” and records all three issues as resolved by April 20 in v2.1.116. The reasoning-history bug also produced cache misses, which Anthropic believes contributed to faster usage-limit consumption; a broader evaluation later found a 3% drop associated with the verbosity prompt for Opus 4.6 and 4.7. The engineering consequence is that context management, effort defaults, and prompt edits all belong in quality evaluation, with public-build testing and per-model checks, even when the underlying inference layer remains unchanged. [Original article](https://www.anthropic.com/engineering/april-23-postmortem)

Anthropic Engineering 的《An update on recent Claude Code quality reports》是一篇四月的故障复盘，可作为理解当前 agent 工具环境讨论的历史背景。文章将质量问题归因于三项产品改动：降低默认推理强度；会话闲置超过一小时后，缓存逻辑中的缺陷反复移除先前的推理内容；以及一项旨在减少冗长输出的系统提示词改动。文章明确表示“API 未受影响”，并记录三项问题已在 4 月 20 日的 v2.1.116 中全部解决。反复丢失推理历史还造成了缓存未命中，Anthropic 认为这促使使用额度消耗加快；后续更广泛的评估发现，减少输出的提示词让 Opus 4.6 和 4.7 在其中一项评估中下降了 3%。这说明，即使底层推理服务没有变化，上下文管理、推理强度默认值和提示词修改仍须纳入质量评估，并通过公开版本测试及针对各模型的检查来验证。 [原文](https://www.anthropic.com/engineering/april-23-postmortem)