九月十二日

OpenAI’s Thibault Sottiaux reports fixes for Astra quality regressions involving over-triggered legacy skills, an opt-in context-management experiment affecting an estimated 4,000–5,000 users, and misconfigured engines. He says the experiment was disabled and the faulty engines removed, improving follow-through and tracking of the latest message. OpenAI has also welcomed Aidan and Sasha from Git AI and will keep investing in its open-source coding-agent contribution tool. Sottiaux lists Images 2.5, GPT-Live-1, Agents API, Data Agent, and ChatGPT for Financial Services among the week’s Astra-powered releases.

OpenAI 的 Thibault Sottiaux 表示，团队修复了多项 Astra 质量问题，涉及过度触发的旧版 skills、一项估计影响 4,000 至 5,000 名用户的自愿参与式上下文管理实验，以及配置不当的引擎。他说，实验已停用，有问题的引擎已移除，任务执行的完整性和对最新消息的跟踪因此得到改善。OpenAI 还迎来了 Git AI 团队的 Aidan 和 Sasha，并将继续投入其用于分析 coding agent 代码贡献的开源工具。Sottiaux 列出的本周 Astra 驱动产品包括 Images 2.5、GPT-Live-1、Agents API、Data Agent 和 ChatGPT for Financial Services。

Anthropic Claude Code builder Thariq says plugin evals can now help check whether skills still work after model releases; initialize them with `claude plugin eval init` in the plugin folder. He also cautions against reading pass/fail scores in isolation: overly strict hidden tests can reject answers that make more sense than the expected result.

Anthropic Claude Code 开发者 Thariq 表示，plugin evals 现已可以帮助检验 skills 在模型更新后是否仍然有效，在插件目录运行 `claude plugin eval init` 即可初始化。他还提醒，不应孤立地看待通过或失败的分数：过于严格的隐藏测试可能否定实际上比预期答案更合理的回答。

Every CEO Dan Shipper argues that higher benchmark scores reveal little about performance on an individual’s real work. His team built an internal platform for employees to create personal benchmarks from daily tasks, adding quantitative checks to its hands-on model reviews.

Every CEO Dan Shipper 认为，更高的 benchmark 分数并不能充分说明模型处理个人实际工作的表现。团队搭建了内部平台，让员工用日常任务创建个人 benchmark，为基于实际使用的模型评测加入量化检查。

Meta Senior Director of AI Madhu Guru attributes enterprise AI failures to outdated product playbooks, weak investment in evals, and central platform teams disconnected from the work they aim to improve. He recommends hiring experienced AI product leaders, making evals a core capability, and embedding builders directly in finance, sales, and support teams.

Meta AI 高级总监 Madhu Guru 将企业 AI 项目的失败归因于过时的产品开发方法、对 evals 投入不足，以及中央平台团队脱离了它们试图改善的实际工作。他建议聘用有 AI 产品经验的负责人，把 evals 作为核心能力，并让开发者直接进入财务、销售和客服团队共同建设产品。

Box CEO Aaron Levie says Box can now be mounted into agent sandboxes, allowing agents to read and write files on their computers. He argues that agents executing critical enterprise workflows need the same basic capabilities that people have long relied on.

Box CEO Aaron Levie 表示，现在可以将 Box 挂载到 agent sandbox，让 agent 在其计算机上读写文件。他认为，执行关键企业工作流的 agent 需要人类长期依赖的那些基础能力。

Vercel CEO Guillermo Rauch says Tailscale’s model router runs on Vercel AI Gateway. He compares AI gateways to CDNs, arguing that direct connections to model providers are brittle and that building routing infrastructure independently is costly and cumbersome.

Vercel CEO Guillermo Rauch 表示，Tailscale 的模型路由器以 Vercel AI Gateway 为底层基础设施。他将 AI gateway 比作 CDN，认为直接连接模型提供商的方案较脆弱，而自行建设路由基础设施又昂贵且繁琐。

AI educator Peter Yang is skeptical that “software factories” can build new features or improve products end to end without human requirements and review, except in verification and testing. In his experience, a single mistaken assumption can waste an overnight run. Separately, he is consolidating local scheduled tasks in Codex and moving cloud tasks to Grok Bot.

AI 教育创作者 Peter Yang 怀疑，除验证和测试外，“软件工厂”目前能否在没有人类定义需求和审查的情况下，端到端构建新功能或改进产品。按他的经验，一个错误假设就可能浪费整晚运行。他还表示，正在把本地定时任务集中到 Codex，并将云端任务迁往 Grok Bot。

Designer Ryo Lu, whose background includes Cursor, Notion, and Stripe, announces long-lived agents in Cursor for larger ideas. His post links to the announcement without specifying runtime limits or implementation details.

曾参与 Cursor、Notion 和 Stripe 设计的 Ryo Lu 宣布，Cursor 已加入面向大型创意的 long-lived agents。帖子附有发布链接，但未说明运行时限或实现细节。

Replit CEO Amjad Masad says Replit acquired a business built entirely on Replit. He expects more such acquisitions, though the post does not name the business or disclose deal terms.

Replit CEO Amjad Masad 表示，Replit 收购了一家完全使用 Replit 构建的企业。他预计未来会有更多此类收购，但帖子没有披露企业名称或交易条款。

OpenClaw and OpenAI builder Peter Steinberger demonstrates Astra playing Doom in an OpenClaw cloud session using CUA. He also reports submitting a patch to trycua to make keyboard input reliable under Linux.

OpenClaw 和 OpenAI 开发者 Peter Steinberger 展示了 Astra 在 OpenClaw 云端会话中通过 CUA 玩 Doom。他还表示，自己向 trycua 提交了补丁，以提高 Linux 下键盘输入的可靠性。

Builder Zara Zhang calls the “one-person company” overrated even as AI lets individuals accomplish more. Her concern is motivation: building alone removes the partner who helps brainstorm, endure setbacks, and celebrate progress.

开发者 Zara Zhang 认为，即便 AI 让个人能完成更多工作，“一人公司”仍被高估了。她担心的是动力问题：独自创业意味着缺少能够一起讨论想法、承受挫折并庆祝进展的伙伴。

The Takeaway: Coinbase CEO and cofounder Brian Armstrong sees useful agents needing both spending accounts and institutional memory that improves with human feedback.

核心观点： Coinbase CEO 兼联合创始人 Brian Armstrong 认为，要让 agent 真正发挥作用，既要提供支出账户，也要让组织经验随着人类反馈不断积累。

Armstrong’s plan connects an “Everything Exchange” spanning asset classes, global stablecoin payments, and agentic finance. The practical bottleneck is small payments: he says roughly 76% of the agent-commerce transactions Coinbase observes are under 30 cents, making a typical 30-cent fixed card fee uneconomic. “We don't want the AIs to be unbanked.” He describes self-custodial wallets and plans for segregated agent accounts linked to a human identity, with x402 supporting purchases of data and services. The point is to let agents complete work without stopping whenever they encounter a payment screen.

Armstrong 的计划将覆盖多类资产的“Everything Exchange”、全球 stablecoin 支付和 agentic finance 联系起来。实际瓶颈在于小额支付：他表示，Coinbase 观察到的 agent 商务交易中，约 76% 的金额低于 30 美分，典型的 30 美分银行卡固定手续费因此并不划算。“我们不希望 AI 无法获得金融服务。”他介绍了自托管钱包，以及与人类身份关联、资金隔离的 agent 账户计划，并提到 x402 支持数据和服务采购。目的在于让 agent 完成任务，而不是每遇到支付页面就停下来。

Inside Coinbase, the more concrete improvement loop starts before an agent changes a service. It reads a team or repository “brain” containing incident history, financial controls, experiments, and accepted or rejected pull requests. When a reviewer corrects the result, that reasoning must go back into the shared context. Armstrong says this helps first-attempt PR acceptance improve over time; the mechanism depends on human review becoming reusable memory. Coinbase’s internal Toshi harness also integrates tools and can pay external vendors through x402.

在 Coinbase 内部，更具体的改进循环从 agent 修改服务之前就开始了。它会读取团队或代码库的“brain”，其中保存事故历史、财务控制要求、实验，以及被接受或拒绝的 pull requests。审查者修正结果时，修正背后的理由必须写回共享上下文。Armstrong 表示，这有助于逐步提高 PR 一次通过率；其机制依赖于把人工审查转化为可复用的记忆。Coinbase 的内部 Toshi harness 还整合了工具，并能通过 x402 向外部供应商付款。

His other company, New Limit, applies AI to choosing experiments for cellular reprogramming, then tests candidates in the lab. Across both efforts, the common pattern is a feedback loop: computational proposals become more useful when grounded in accumulated operational or experimental results.

他参与创办的另一家公司 New Limit 用 AI 选择细胞重编程实验，再在实验室中测试候选方案。这两项工作的共同点是反馈循环：计算产生的方案，需要依靠持续积累的实际运行或实验结果来提高价值。

通过 Follow Builders skill 生成：https://github.com/zarazhangrui/follow-builders