AI Builders Digest — 2026-09-18

2026-09-18

AI Builders Digest — 2026-09-18

X / TWITTER

Boris Cherny, Claude Code at Anthropic

Claude is collapsing chat, Cowork, Docs, Slides, and Design into one context-aware interface. The important product shift is automatic routing: users state the outcome, while Claude decides whether the task needs a quick answer, deeper agentic work, or an editable artifact. This moves AI products away from tool selection and toward intent-driven orchestration.

Claude 正在把 chat、Cowork、Docs、Slides 和 Design 合并为一个共享上下文的界面。真正重要的产品变化是自动路由：用户只需描述结果，Claude 自行判断应快速回答、执行更深层的 agent 工作，还是生成可编辑文件。这意味着 AI 产品正在从“让用户选择工具”转向“围绕意图自动编排”。

Cat Wu, Claude Code and Cowork at Anthropic

Cat Wu explains why Anthropic merged Cowork into Claude chat: users did not want to decide which Claude product matched each task. Claude now chooses the effort level and output type from the prompt, while preserving user control to stop or redirect the work. The rollout will be gradual as Anthropic tunes speed and reliability.

Cat Wu 解释了 Anthropic 合并 Cowork 与 Claude chat 的原因：用户不想先判断某项任务应该使用哪个 Claude 产品。现在 Claude 会根据 prompt 自动选择投入程度和输出类型，同时保留用户随时停止或调整方向的控制权。Anthropic 将逐步推出该体验，以继续优化速度与可靠性。

Claude, Anthropic's AI assistant

Claude can now create decks, documents, and designs directly inside a conversation, with manual editing, comments, presentation, export, and link sharing built into the same flow. Existing Cowork chats, projects, artifacts, connectors, and skills remain available, reducing migration friction for current users.

Claude 现在可以直接在对话中创建演示文稿、文档和设计，并在同一流程中支持手动编辑、评论、演示、导出和链接分享。原有 Cowork 的聊天、项目、artifacts、connectors 与 skills 都会保留，从而降低现有用户的迁移成本。

Thariq, Claude Code at Anthropic

Thariq updates his view on agent tooling: bash alone is no longer sufficient for reliable tool calling, although sandboxed bash remains valuable for code generation and execution. He favors giving models purpose-shaped tools, such as a database API instead of a filesystem abstraction, and sees Claude Managed Agents as a useful balance because the sandbox can remain optional and independent from the agent loop.

Thariq 更新了他对 agent 工具设计的判断：仅靠 bash 已不足以实现可靠的 tool calling，但沙箱化 bash 对代码生成与执行仍然有价值。他更倾向于给模型提供贴近真实目的的工具，例如直接提供数据库 API，而不是用文件系统做间接抽象；Claude Managed Agents 的优势则是沙箱可选，并能与 agent loop 独立运行。

Guillermo Rauch, Vercel CEO

Vercel CEO Guillermo Rauch reports that typesafeai's Jev model is up to 18 times faster at p95 than the GPT Luna safety reviewer currently used by fx, while also being more accurate. Vercel plans to bring it to AI Gateway and is considering it as the new default, showing how narrow, fast classifiers can become critical control-plane components in agent systems.

Vercel CEO Guillermo Rauch 表示，typesafeai 的 Jev 模型在 p95 延迟上比 fx 当前使用的 GPT Luna 安全审查器快最多 18 倍，同时准确率更高。Vercel 计划将其接入 AI Gateway，并可能设为新默认模型。这说明轻量、快速的专用分类器正在成为 agent 系统控制平面的关键组件。

Aaron Levie, Box CEO

Box CEO Aaron Levie argues that very fast, low-cost models unlock enterprise agent workflows that larger models make uneconomical. Classification, routing, domain-specific decisions, and safety checks are not peripheral tasks; they are gates repeated throughout business processes, making latency and unit cost strategically important.

Box CEO Aaron Levie 认为，超快、低成本模型将解锁许多用大模型难以经济运行的企业 agent 工作流。分类、路由、领域决策与安全检查并非边缘任务，而是业务流程中反复出现的关键门槛，因此延迟和单位成本具有战略意义。

Peter Yang, AI educator and podcaster

Peter Yang describes an eight-skill production system for his podcast. Dedicated skills handle guest research, interview preparation, transcript editing, quote selection, and orchestration into six derivative assets. His practical conclusion is that stronger foundation models do not eliminate skills; skills remain useful for encoding repeatable editorial judgment and browser workflows.

AI 教育者与播客创作者 Peter Yang 分享了一套由八个 skills 组成的播客生产系统，分别负责嘉宾研究、采访准备、逐字稿编辑、金句筛选，以及编排生成六类衍生内容。他的实践结论是：更强的基础模型并不会让 skills 失去价值，skills 仍适合固化可重复的编辑判断和浏览器工作流。

Madhu Guru, Senior Director of AI at Meta

Meta AI Senior Director Madhu Guru argues that safety and security should be product and model features, not external guardrails added after the system is built. For builders, this reframes safety from a compliance layer into an architectural requirement.

Meta AI 高级总监 Madhu Guru 认为，安全与防护应该成为产品和模型本身的功能，而不是系统完成后再从外部附加的 guardrails。对 builder 而言，这意味着应把安全从合规层重新定义为架构层要求。

Nikunj Kothari, FPV Ventures partner

FPV Ventures partner Nikunj Kothari has been testing a shared household agent built with Nous Research technology. He highlights granular permissions and execution capabilities, including reading only selected emails, converting inline attachments into structured data, and using authenticated browser sessions. The notable product pattern is a multi-user agent that becomes useful through tightly scoped access rather than broad autonomy.

FPV Ventures 合伙人 Nikunj Kothari 正在测试一个基于 Nous Research 技术构建的家庭共享 agent。他强调了细粒度权限和执行能力，例如只读取指定邮件、把内嵌附件转成结构化数据，以及使用已登录的浏览器会话。值得关注的产品模式是：多用户 agent 的实用性来自严格限定的访问权限，而不是无限扩大自主权。

Zara Zhang, builder

Builder Zara Zhang says she is using Claude less because its communication style feels performatively sophisticated instead of direct. The high engagement around the complaint is a product signal: model tone and communicative efficiency can drive retention just as materially as benchmark capability.

Builder Zara Zhang 表示，她最近减少使用 Claude，因为其表达方式显得刻意展示复杂与聪明，而不是直接传递观点。该帖获得大量互动，本身就是一个产品信号：模型语气和沟通效率对留存的影响，可能与 benchmark 能力同样显著。

Garry Tan, Y Combinator President and CEO

Y Combinator President and CEO Garry Tan points toward a personal-agent architecture where the same personality and full memory can move across different harnesses. This separates the durable user relationship and memory layer from the execution environment, making the harness replaceable rather than the center of the product.

Y Combinator 总裁兼 CEO Garry Tan 指向一种个人 agent 架构：相同的人格与完整记忆可以跨不同 harness 使用。这将持久的用户关系和记忆层与执行环境解耦，使 harness 成为可替换组件，而不再是产品中心。

Sam Altman, OpenAI CEO

OpenAI CEO Sam Altman says the main launch he expected this week has moved to next week and will be worth the wait. No product details were provided, so the signal is timing only rather than a confirmed capability announcement.

OpenAI CEO Sam Altman 表示，原计划本周发布的主要产品将推迟到下周，但“值得等待”。由于没有提供产品细节，目前能确认的只有发布时间变化，而非具体能力。

Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders