九月十九日

Box CEO Aaron Levie demonstrated Jev reading an incident report, judging customer impact and severity, routing the file into an escalate, monitor, or review folder, and recording the result as metadata. He describes the workflow as nearly instantaneous and almost free, and sees similar uses in insurance claims, contracts, loans, security reviews, and customer logs.

Box CEO Aaron Levie 演示了 Jev 如何读取事故报告，判断客户影响和严重程度，再将文件移入 escalate、monitor 或 review 文件夹，并把结果记录为 metadata。他表示，这套流程几乎瞬间完成、成本接近零，也可用于保险理赔、合同、贷款、安全审查和客户日志等场景。

[Source](https://x.com/levie/status/2101007708044574906)

FPV Ventures partner Nikunj Kothari shared a Jev demo that scored 3,000 children's snacks against multiple criteria in 28 seconds for $0.11. He also built Jevable, a collection of community demos that can be filtered by category.

FPV Ventures 合伙人 Nikunj Kothari 分享了一项 Jev 演示：在 28 秒内按多个标准评估 3,000 种儿童零食，花费 0.11 美元。他还创建了 Jevable，按类别整理社区演示，方便查找不同用法。

[Scoring source](https://x.com/nikunj/status/2101006585481073093) · [Snack catalog](https://t.co/KSiodPdZHT) · [Demo link 1](https://t.co/c9O6uTyiUK) · [Demo link 2](https://t.co/7Su6QACPY4) · [Jevable source](https://x.com/nikunj/status/2101077053567332618) · [Jevable](https://t.co/BImN16sFqz)

Vercel CEO Guillermo Rauch reported that open models accounted for 78.4% of token volume on Vercel AI Gateway, versus 21.6% for closed models, potentially a record share. Moonshot AI and DeepSeek ranked third and fourth by spend; together with Z.ai, their inference spend exceeded second-place OpenAI. He cautioned that this measures spending across inference providers, mostly in the United States, rather than revenue paid directly to open-weight labs. In a separate post about Jev adoption, he connected demand to frustration with expensive, slow AI and developers' interest in using AI in more places.

Vercel CEO Guillermo Rauch 表示，open models 占 Vercel AI Gateway token 用量的 78.4%，closed models 占 21.6%，这一占比可能创下纪录。按支出计算，Moonshot AI 和 DeepSeek 分列第三、第四；加上 Z.ai 后，三者的 inference 支出合计超过排名第二的 OpenAI。他提醒，这里统计的是跨 inference provider 的支出，主要发生在美国，并不等于直接流向 open-weight 实验室的收入。他在另一条关于 Jev 采用情况的帖子中，将需求归因于开发者对 AI 成本和速度的不满，以及把 AI 用到更多环节的意愿。

[Gateway source](https://x.com/rauchg/status/2101186741042663579) · [Adoption source](https://x.com/rauchg/status/2101079472732848510) · [Related link 1](https://t.co/tK86PNGgNs) · [Related link 2](https://t.co/Y7w3UKosBC)

Anthropic’s Claude Code builder Thariq announced that version 2.1.277 checks for `AGENTS.md` when a folder has no `CLAUDE.md`; users can toggle the behavior in `/config`. The support is implemented as a built-in Claude Code mod, previewing an upcoming customization mechanism that will let developers create their own versions of project-instruction handling. Thariq also shared the mod’s source.

Anthropic 的 Claude Code 开发者 Thariq 宣布，从 2.1.277 版本开始，如果文件夹中没有 `CLAUDE.md`，Claude Code 会继续查找并使用 `AGENTS.md`；用户可以在 `/config` 中开关这一行为。这项支持通过内置 Claude Code mod 实现，也预示着即将推出的自定义机制，开发者之后可以自行构建处理项目指令的方式。Thariq 还公开了该 mod 的源码。

[Announcement](https://x.com/trq212/status/2101009392611278961) · [Implementation](https://x.com/trq212/status/2101009393731223817) · [Source post](https://x.com/trq212/status/2101009395052343462) · [Mod source](https://t.co/ZGJxRbh33b)

AI tutorial creator Peter Yang called Meta’s Muse the best personal agent he has tried and said it saved him more than $800 per year on cable and phone bills. His examples include a personalized morning news feed, a habit tracker, and having the agent call customer support; one transcript showed Muse negotiating $288 in annual Comcast savings. Yang’s practical warning is that many companies’ support lines are not ready for agents acting on behalf of customers.

AI 教程创作者 Peter Yang 称 Meta 的 Muse 是他目前体验过最好的 personal agent，并表示它每年为自己的有线电视和手机账单节省了 800 多美元。他列出的用例包括个性化晨间新闻、习惯追踪，以及让 agent 致电客服；其中一份通话记录显示，Muse 从 Comcast 谈下了每年 288 美元的费用减免。Yang 给出的现实提醒是，许多公司的客服热线尚未准备好应对代表客户行动的 agent。

[Overview source](https://x.com/petergyang/status/2101033599319613533) · [Transcript source](https://x.com/petergyang/status/2101083891507593576) · [Video](https://t.co/jte6evu5iI)

OpenClaw and OpenAI builder Peter Steinberger described roboclaw running his team's server, participating in Discord through gpt-live, and tracking the sessions it manages. The team can ask it for context from current and past sessions during meetings. He also suggested using the home sidebar to ask an agent to reorganize sessions.

OpenClaw 与 OpenAI 开发者 Peter Steinberger 表示，roboclaw 正在运行团队服务器，通过 gpt-live 活跃于 Discord，并掌握自己管理的多个 session。团队开会时，可以直接询问当前和历史 session 的上下文。他还建议从 home sidebar 让 agent 帮忙重新整理 session。

[Team server source](https://x.com/steipete/status/2101141707375227372) · [Session organization source](https://x.com/steipete/status/2101139037801283997) · [Related link](https://t.co/BFw01me3sP)