Your agent made the right call — and your system still broke. The user clicked "confirm refund", the page refreshed, the approval banner vanished, and nobody knows whether the refund should run. Worse: the request reached the payment gateway, the backend restarted before writing local state, and recovery re-triggered the refund. Twice. That's the gap between "an agent can run one loop" an …
模型判断对了，系统照样会出事。用户点了「确认退款」，页面突然刷新，审批条消失——前端不知道审批还生不生效，后端不知道用户到底有没有做决定；更危险的是，退款请求已经提交给支付渠道，后端却在写入本地状态前重启了，恢复后系统只看到「任务尚未完成」，于是又调了一次退款接口。 这就是「能跑一个 ReAct 循环」和「能跑一个产品」之间的差距。本文拆解 vivo KDC 工程系列（文末有链接）的核心做法：把 Agent 的目标、判断、
Ornith 1.5 open weights train on self-generated tasks: Terminal-Bench 2.1 rose from 77.5 to 86.1. Next battleground: the question bank.
Ornith-1.5 开源权重：模型自己出题、自己搭脚手架、自己跑轨迹，两月自测 Terminal-Bench 2.1 从 77.5 涨到 86.1。下一场开源与闭源的竞赛，战场在题库。
China's model race took a fresh turn on Thursday: Zhipu AI confirmed that the viral "Niu Lai" (bull market) model is its own GLM-5.3 Flash — the lab's first natively multimodal model, trained on domestic chips. Meanwhile, video-generation apps are turning into real revenue lines, Alibaba quietly open-sourced another Qwen model, and MiniMax crossed the $800M annualized-revenue mark. Here's what mat
每日精选全球 AI 资讯，快速掌握行业趋势 🔥 今日亮点（Top 3） 1. 神秘「牛来」模型果然是智谱！GLM首个原生多模态，还用的国产卡 核心洞察：国产模型快速崛起 来源：https://www.qbitai.com/2026/08/479919.html 2. AI视频应用井喷，美图打开新的增长空间 核心洞察：AI 行业持续火热 来源：https://www.qbitai.com/2026/08/479895.html 3. 小宇宙推出《AI趋势报告》：AI创作、AI办公、协作型AI等成讨论新趋势 核心洞察：新产品密
Veeda AI builds a multimodal world model where robots learn by trial and error in scalable simulations, avoiding costly real-world practice.
OpenAI shared first Jalapeño benchmark results. The custom inference chip outperformed Nvidia Blackwell on InferenceX throughput and latency.
OpenAI 首款自研推理芯片 Jalapeño 的首批基准测试结果显示，在 SemiAnalysis 的 InferenceX 上，其每瓦算力与端到端延迟均优于英伟达 Blackwell 系统，意味着领先 AI 实验室正通过自研芯片降低推理成本并重新掌控基础设施。
Shopify CEO Tobi Lütke says he may ban Claude Code unless Anthropic supports AGENTS.md, turning a single Markdown file into the front line of the coding-agent standards war.
Shopify CEO Tobi Lütke 公开要求 Anthropic 让 Claude Code 原生支持 AGENTS.md，否则将在全公司禁用。这一事件把一份 Markdown 文件推上了编码 Agent 标准战争的正面战场。