AIE 上海 2026 on Yousa Driven Development | YDDhttps://yousali.com/events/aie-shanghai-2026/Recent content in AIE 上海 2026 on Yousa Driven Development | YDDHugo -- 0.148.1zhTue, 10 Nov 2026 09:05:10 +0800Reading the Web at Trillion-Token Scalehttps://yousali.com/events/aie-shanghai-2026/s37-jina-reading-the-web/Fri, 06 Nov 2026 17:20:00 +0800https://yousali.com/events/aie-shanghai-2026/s37-jina-reading-the-web/AI 读网页的 token 成本已和广告千次浏览收入相当，网页转 Markdown 后还要再压缩；讲者不看好 agent 按请求付费，更想让 agent 看广告或做可验证任务换免费访问。Memory Engineering, RSI and Team Knowledge Layerhttps://yousali.com/events/aie-shanghai-2026/s36-nowledge-memory-engineering/Fri, 06 Nov 2026 17:10:00 +0800https://yousali.com/events/aie-shanghai-2026/s36-nowledge-memory-engineering/agent memory 要管记什么、时间、脉络和遗忘；今年 4 月起半闭环的自我改进在团队日常里跑通，卡点在人的验证，所以团队知识宜做成 harness 能直接用的上下文层，而不是训进参数。Robot Data Needs a Build Systemhttps://yousali.com/events/aie-shanghai-2026/s35-lancedb-robot-data-build-system/Fri, 06 Nov 2026 17:00:00 +0800https://yousali.com/events/aie-shanghai-2026/s35-lancedb-robot-data-build-system/机器人数据采集一次却要反复重新加工，需要像代码构建系统那样按 cell 标脏、增量重算、按快照发布；LanceDB 从文件格式、表格式到计算层一路往上做这件事。From Hardware to Intelligence: Home Service Robotshttps://yousali.com/events/aie-shanghai-2026/s34-futuring-robot-home-service-robots/Fri, 06 Nov 2026 16:40:00 +0800https://yousali.com/events/aie-shanghai-2026/s34-futuring-robot-home-service-robots/家庭服务机器人靠世界动作模型加真实执行数据自动回流训练来提升成功率，前提是软硬件时间对齐，燃料来自把机器人长期放进真实用户家里。Should This Agent Be Allowed to Do That?https://yousali.com/events/aie-shanghai-2026/s33-1password-agent-intent-check/Fri, 06 Nov 2026 16:25:00 +0800https://yousali.com/events/aie-shanghai-2026/s33-1password-agent-intent-check/agent 的意图落在每次工具调用上，应在执行前用本地、低延迟的意图裁判按用户任务判定；她的实验里约 4 亿参数的判别类小模型明显准于生成式小模型，拿不准时转人工。Building Robot Hardware for Rapidly Evolving Modelshttps://yousali.com/events/aie-shanghai-2026/s32-dyna-robot-hardware/Fri, 06 Nov 2026 16:05:00 +0800https://yousali.com/events/aie-shanghai-2026/s32-dyna-robot-hardware/通用机器人硬件跟不上模型每周一次的迭代，所以硬件不该追模型，而应做成可预测的稳定接口；DYNA 新机器人弃谐波、改用行星减速器就是按这条原则选的。Noe-0: A World Action Model Built on Real-World Human Experiencehttps://yousali.com/events/aie-shanghai-2026/s31-noematrix-noe-0/Fri, 06 Nov 2026 15:50:00 +0800https://yousali.com/events/aie-shanghai-2026/s31-noematrix-noe-0/具身智能落地的关键是获取新能力的成本；穹彻用无本体采集的大规模真实数据训练世界模型，加力觉和 agentic 系统，把新任务交付从大团队几个月压向一人两周。Real2Real: A Quantum Embodied AI Paradigm for Optimal Behaviors Generationhttps://yousali.com/events/aie-shanghai-2026/s30-quantsparkle-real2real/Fri, 06 Nov 2026 15:30:00 +0800https://yousali.com/events/aie-shanghai-2026/s30-quantsparkle-real2real/把动作分布建成玻尔兹曼能量地形，用光量子相干伊辛机在毫秒级采全局最低点，再串成只靠真实数据的自我改进循环；讲者称因此突破了 GPU 训练模型的成功率平台期。Communication Among Sovereign Agents: The Foundation of Agent Economyhttps://yousali.com/events/aie-shanghai-2026/s29-masons-sovereign-agents/Fri, 06 Nov 2026 15:05:00 +0800https://yousali.com/events/aie-shanghai-2026/s29-masons-sovereign-agents/agent 的关键在行动，行动要改别人的环境；品牌不愿被降格成 API，所以 agent 经济需要一层让保有主权的 agent 互相通信、交换意图的基础层。How We're Building Open Decision Models for On-Device Agentshttps://yousali.com/events/aie-shanghai-2026/s28-liquid-ai-decision-models/Fri, 06 Nov 2026 14:50:00 +0800https://yousali.com/events/aie-shanghai-2026/s28-liquid-ai-decision-models/agent 里很多 LLM 调用只是在做小决策，不必逐 token 生成；决策模型一次前向给出答案和置信度，用 LoRA 微调、双向注意力、温度缩放校准，就能做成跑在手机和笔记本上的小模型。From 3D Generation to World Models: Building Persistent Interactive Worlds for Physical AIhttps://yousali.com/events/aie-shanghai-2026/s27-vast-3d-to-world-models/Fri, 06 Nov 2026 14:35:00 +0800https://yousali.com/events/aie-shanghai-2026/s27-vast-3d-to-world-models/coding agent 已能用代码搭 3D 场景，Tripo 的答案是分工：说得清的交给 agent 写代码，只能认出来的细节交给模型学；再用程序管状态、模型管画面，做出改动能留下来的可交互世界。Beyond Vision-Language-Action: First Principles for Building Physical AIhttps://yousali.com/events/aie-shanghai-2026/s26-aimoga-beyond-vla/Fri, 06 Nov 2026 14:15:00 +0800https://yousali.com/events/aie-shanghai-2026/s26-aimoga-beyond-vla/工业现场用不上纯 VLA，也养不起传统视觉模型，所以要在 VLA 外面包一层感知动作结果、发现偏差、修正并沉淀经验的闭环，安全监督另走不依赖模型的硬件逻辑。Stronger Models, Leaner Harnesses: What to Control and What to Let Gohttps://yousali.com/events/aie-shanghai-2026/s25-minimax-leaner-harnesses/Fri, 06 Nov 2026 13:55:00 +0800https://yousali.com/events/aie-shanghai-2026/s25-minimax-leaner-harnesses/模型变强以后，harness 里的提示词、subagent、todo/plan、压缩策略都该拿评测逐项检验，能删就删；不变的追求只有两条：更少的 token、更高的任务完成率。圆桌：AI startups: products, markets & what comes nexthttps://yousali.com/events/aie-shanghai-2026/s24-panel-ai-startups-vc/Fri, 06 Nov 2026 13:15:00 +0800https://yousali.com/events/aie-shanghai-2026/s24-panel-ai-startups-vc/三位投资人都认为 AI 应用还没真正爆发，帮人省时间的 personal agent 天花板有限，情绪价值这类能扩张的需求更值得做；在选人上，三人对教授创业、年轻团队和 RSI 方向看法不一。The Good, the Bad and the Ugly: What 3,000 live sessions taught us about streaming speech-to-texthttps://yousali.com/events/aie-shanghai-2026/s23-nus-streaming-stt/Fri, 06 Nov 2026 11:40:00 +0800https://yousali.com/events/aie-shanghai-2026/s23-nus-streaming-stt/流式语音 agent 转写出没人说过的字，主要不怪 STT 模型，而是 VAD 把音频切得太碎、又没告诉模型语言；指定语言、等更长停顿再切，能找回大部分准确率。Trusting Evals I Can't Readhttps://yousali.com/events/aie-shanghai-2026/s22-daily-trusting-evals-i-cant-read/Fri, 06 Nov 2026 11:25:00 +0800https://yousali.com/events/aie-shanghai-2026/s22-daily-trusting-evals-i-cant-read/写东西变便宜后谁都读不过来，开发者要从检查产出改为检查「检查者」：英文写验收标准，LLM 写代码、当裁判、扮用户，人只审裁判。AI in Manufacturing: The Last Mile Is Half the Journeyhttps://yousali.com/events/aie-shanghai-2026/s21-hicap-ai-in-manufacturing/Fri, 06 Nov 2026 11:10:00 +0800https://yousali.com/events/aie-shanghai-2026/s21-hicap-ai-in-manufacturing/制造业 AI 落地难在模型之前：车间的事大多没变成数据、口径不统一、老设备接不上、工人怕被替代；先做见效快的项目，再梳理计划、语义和流程。AGI is decoding. Encoding is the moat.https://yousali.com/events/aie-shanghai-2026/s20-chiling-encoding-is-the-moat/Fri, 06 Nov 2026 10:45:00 +0800https://yousali.com/events/aie-shanghai-2026/s20-chiling-encoding-is-the-moat/AGI 是最强的解码器，编码清晰的领域终会被它吞没；垂直 AI 的壁垒在于把隐性知识多、交易成本高的实体行业数字化，讲者选了美妆，从消费者端切入。HTML is all your agents need: Redefining Context Engineeringhttps://yousali.com/events/aie-shanghai-2026/s19-heygen-html-is-all-your-agents-need/Fri, 06 Nov 2026 10:25:00 +0800https://yousali.com/events/aie-shanghai-2026/s19-heygen-html-is-all-your-agents-need/上下文的「形状」比记忆、工具、RAG 更根本；HeyGen 把视频编辑器底下的私有 JSON 换成让 agent 直接写 HTML/CSS/JS，模型变强产品就跟着变强。Taxonomy of World Modelshttps://yousali.com/events/aie-shanghai-2026/s18-world-labs-taxonomy-of-world-models/Fri, 06 Nov 2026 10:05:00 +0800https://yousali.com/events/aie-shanghai-2026/s18-world-labs-taxonomy-of-world-models/「世界模型」一词已被用滥，应按功能、领域、尺度与分辨率、覆盖度来归类；讲者认为世界模型会成为 agent 抽象推理与感知/行动之间的接口。Life of a Game Developer in AI erahttps://yousali.com/events/aie-shanghai-2026/s17-kimi-game-dev-in-ai-era/Thu, 05 Nov 2026 17:35:00 +0800https://yousali.com/events/aie-shanghai-2026/s17-kimi-game-dev-in-ai-era/AI 能独立跑长程任务以后，做一个可玩游戏原型的时间和成本降到很低，开发者的角色从动手写变成定目标、验收和做决定。Reinforcement Learning at Scale on Agent Substrate: Orchestrating Millions of Sandboxeshttps://yousali.com/events/aie-shanghai-2026/s16-google-rl-sandboxes-at-scale/Thu, 05 Nov 2026 17:25:00 +0800https://yousali.com/events/aie-shanghai-2026/s16-google-rl-sandboxes-at-scale/前沿实验室的 RL 后训练被大规模沙箱调度卡住；Google 开源了一套 RL rollout 沙箱基础设施，能做到百万级并发、首条指令 1 秒内，下一代用快照省掉沙箱空等时间。Optimizing Organization-Wide LLM Inference Cost: Seat License, Pay-per-Token, or Provisioned, and Beyondhttps://yousali.com/events/aie-shanghai-2026/s15-techcombank-inference-cost/Thu, 05 Nov 2026 17:10:00 +0800https://yousali.com/events/aie-shanghai-2026/s15-techcombank-inference-cost/按「每百万 token 有效成本」算出月用量临界点（他们的例子约 830 亿 token），超过就从按 token 付费转自租 GPU 部署，预计省 30–50%，但要留按 token 付费兜底。From Matching to Delivery: Scaling Creator Partnerships with AIhttps://yousali.com/events/aie-shanghai-2026/b12-ahacreator-matching-to-delivery/Thu, 05 Nov 2026 16:45:00 +0800https://yousali.com/events/aie-shanghai-2026/b12-ahacreator-matching-to-delivery/海外达人营销还停在邮件加表格的手工阶段，传统代理商做 KOC/UGC 不赚钱；AhaCreator 用带真实订单找达人的双边平台，把匹配到数据回收的大部分环节交给 AI。Nobody Planned AI Infra: How AI Builders Pushed a Database Beyond the Databasehttps://yousali.com/events/aie-shanghai-2026/s14-tidb-nobody-planned-ai-infra/Thu, 05 Nov 2026 16:40:00 +0800https://yousali.com/events/aie-shanghai-2026/s14-tidb-nobody-planned-ai-infra/顺着一个用户请求，agent 的进度、文件、应用数据库、trace 分析一步步压到数据层上；TiDB 没有事先计划做这些，是用户需求推着它超出了数据库的范围。From Idea to Viral Launch: How I Built a 3D Little Prince Game with AIhttps://yousali.com/events/aie-shanghai-2026/b11-sac-idea-to-viral-launch/Thu, 05 Nov 2026 16:35:00 +0800https://yousali.com/events/aie-shanghai-2026/b11-sac-idea-to-viral-launch/一个不懂建模的人 16 天做出 3D《小王子》游戏，靠的是摸清新模型的能力边界、审美和懂社媒；做法是用参考图定目标，让 agent 在 Blender/Unity 里搭场景并截图比对自我迭代。Building a reliable, multi-region LLM gatewayhttps://yousali.com/events/aie-shanghai-2026/m03-optiver-multi-region-llm-gateway/Thu, 05 Nov 2026 16:15:00 +0800https://yousali.com/events/aie-shanghai-2026/m03-optiver-multi-region-llm-gateway/LLM 请求有状态，协议翻译、简单负载均衡、跨区域共享库都会出问题；Optiver 靠会话粘性、功能取交集、备份常态分流、鉴权本地化和全链路 tracing 补，最后放弃协议翻译、自建精简网关。Your Data Agent Needs More Than a Knowledge Graph: From Semantic Layers to Executable Ontologies with Apache Ossiehttps://yousali.com/events/aie-shanghai-2026/b10-datusai-executable-ontologies/Thu, 05 Nov 2026 16:15:00 +0800https://yousali.com/events/aie-shanghai-2026/b10-datusai-executable-ontologies/数据 agent 的瓶颈是数字和决策能不能信；知识图谱只说明事物是什么，要算对数，得在语义层上加本体，并用引擎把规范编译成确定性 SQL、拒绝错误查询。Your Agent's Memory Is Next Quarter's Training Datahttps://yousali.com/events/aie-shanghai-2026/s13-zilliz-memory-is-training-data/Thu, 05 Nov 2026 15:55:00 +0800https://yousali.com/events/aie-shanghai-2026/s13-zilliz-memory-is-training-data/AI 数据每年涨一个数量级，单机优化撑不住，要改成以对象存储为底、热温冷三层的 Lakebase，让成本接近对象存储、服务质量接近数据库，线上服务和离线挖掘共用一份数据。The Harness Is the First Reviewerhttps://yousali.com/events/aie-shanghai-2026/b09-data4life-harness-is-first-reviewer/Thu, 05 Nov 2026 15:50:00 +0800https://yousali.com/events/aie-shanghai-2026/b09-data4life-harness-is-first-reviewer/两个人用 agent 重写十年历史的医疗研究应用 OHDSI Atlas，写出的代码审不完，靠 spec、旧系统对照、CI 关卡和截图检查先审一遍；最关键的是检验 agent 写的测试本身。Trusted Context: Building the Layer Between Enterprise Data and AI Agentshttps://yousali.com/events/aie-shanghai-2026/b08-datahub-trusted-context/Thu, 05 Nov 2026 15:40:00 +0800https://yousali.com/events/aie-shanghai-2026/b08-datahub-trusted-context/企业里 agent 答错多半是缺上下文而非模型不行，所以要在企业数据和 agent 之间建一层上下文层：采集、生成、专家验证，再带着可信度分发给 agent。Powering open source personal agents with Exa searchhttps://yousali.com/events/aie-shanghai-2026/s12-exa-personal-agent-search/Thu, 05 Nov 2026 15:35:00 +0800https://yousali.com/events/aie-shanghai-2026/s12-exa-personal-agent-search/用 mem0、browser-use、OpenRouter 和 Exa 搭一个可检查的开源个人 agent：先用搜索高亮拿最少 token，不够再开浏览器；记忆按人区分，主动提醒要条件全满足才打扰。Your Repo. Your Benchmark.https://yousali.com/events/aie-shanghai-2026/b07-agoda-your-repo-your-benchmark/Thu, 05 Nov 2026 15:30:00 +0800https://yousali.com/events/aie-shanghai-2026/b07-agoda-your-repo-your-benchmark/评估和改进 coding agent 的上下文，要拿自己仓库里已合并的 MR 回放当基准；别人的工具和经验换个仓库结果可能反过来，一次做不成的地方要变成修上下文的 MR。Agentic Harness: Real-Time Audio and Video AI Interviewshttps://yousali.com/events/aie-shanghai-2026/b06-cookiy-agentic-harness/Thu, 05 Nov 2026 15:10:00 +0800https://yousali.com/events/aie-shanghai-2026/b06-cookiy-agentic-harness/AI 用户访谈的成本大头是受访者报酬，所以优化目标是每分钟洞察；他们用三类子 agent 级联、多路并发压延迟、历史坏案例做 benchmark，并坚持不做后训练。AI-Native Work at Manus: One Monorepo, Open Taskshttps://yousali.com/events/aie-shanghai-2026/s11-manus-ai-native-work/Thu, 05 Nov 2026 15:10:00 +0800https://yousali.com/events/aie-shanghai-2026/s11-manus-ai-native-work/Manus 把全公司的代码、知识和内部系统收进一个 Mono 仓库让 AI 看全上下文，再公开每个人的 agent 任务让人看全上下文；AI 强了以后，瓶颈回到人的决策和注意力。Beyond Screenshots: Building Agent-Native Verification for Mobile Appshttps://yousali.com/events/aie-shanghai-2026/b05-meta-beyond-screenshots/Thu, 05 Nov 2026 14:55:00 +0800https://yousali.com/events/aie-shanghai-2026/b05-meta-beyond-screenshots/手机验证智能体只靠截图点屏幕不够用，要让 app 对智能体可观测、提供语义化操作，再用确定性流程加 LLM 搭验证智能体，接进编码内循环和 CI。Kubernetes Is Not Your Sandboxhttps://yousali.com/events/aie-shanghai-2026/s10-daytona-k8s-is-not-your-sandbox/Thu, 05 Nov 2026 14:50:00 +0800https://yousali.com/events/aie-shanghai-2026/s10-daytona-k8s-is-not-your-sandbox/Kubernetes 为长期运行的 Web 服务设计，agent 沙箱要的是短命但有状态；在 K8s 上硬做有状态沙箱会掉进外挂存储的坑，讲者建议交给专门的沙箱服务商。Beyond Reasoning: Building Infrastructure for AI Agents That Acthttps://yousali.com/events/aie-shanghai-2026/b04-visa-infra-for-agents-that-act/Thu, 05 Nov 2026 14:40:00 +0800https://yousali.com/events/aie-shanghai-2026/b04-visa-infra-for-agents-that-act/agent 开始替人花钱，现有支付体系处处不适配；Visa 的做法是发限范围限时效的 token、按用户授权放行、用 Trusted Agent Protocol 让认证过的 agent 与商户握手。On-Device AI: Bringing Intelligence Closer to Everyonehttps://yousali.com/events/aie-shanghai-2026/s09-autoark-on-device-ai/Thu, 05 Nov 2026 14:30:00 +0800https://yousali.com/events/aie-shanghai-2026/s09-autoark-on-device-ai/智能硬件全靠云端 API 会越卖越赔；按设备算力分三档把模型放到端上已经可行，耳机案例里端侧意图分流省掉了 92% 以上的 token 成本。Beyond the Pilot: Managing Your 500th AI Agenthttps://yousali.com/events/aie-shanghai-2026/b03-dify-managing-500th-agent/Thu, 05 Nov 2026 14:20:00 +0800https://yousali.com/events/aie-shanghai-2026/b03-dify-managing-500th-agent/企业 agent 到几百上千个以后，难点从怎么造 agent 变成运营、治理和优化；这个规模来得比人们想的快，要提前规划。Dark Factory to Lobsterville: Intent Engineering for Persistent Multiplayer Agentshttps://yousali.com/events/aie-shanghai-2026/s08-openclaw-intent-engineering/Thu, 05 Nov 2026 14:05:00 +0800https://yousali.com/events/aie-shanghai-2026/s08-openclaw-intent-engineering/用 agent 大规模并行开发 OpenClaw 后，瓶颈从写得多快变成怎么证明改对了；工程师要先定义可验证的结果，再让 agent 爬坡，并把 CI 做便宜做快。Trust Your Own Evals, Not the Leaderboard: What 10k+ Internal Agent PRs Taught Us About Adopting Open Modelshttps://yousali.com/events/aie-shanghai-2026/b02-harvey-trust-your-own-evals/Thu, 05 Nov 2026 14:05:00 +0800https://yousali.com/events/aie-shanghai-2026/b02-harvey-trust-your-own-evals/Harvey 内部 agent 的上万个 PR 都能重放成评测集；用它测几类固定任务后发现，该改的常是数据、小 bug 和评测集本身，选模型要信自己的 eval。Building a Personal Agent for Growth with YouMindhttps://yousali.com/events/aie-shanghai-2026/b01-youmind-personal-agent-for-growth/Thu, 05 Nov 2026 13:45:00 +0800https://yousali.com/events/aie-shanghai-2026/b01-youmind-personal-agent-for-growth/各家 Personal Agent 工程上大同小异，都在向外替人解决烦恼、满足欲望；YouMind 想做向内的事：营造让人和 AI 一起创作、以创作促成长的环境。From Model Intelligence to Agent Harness: Enabling Long-Horizon Agentshttps://yousali.com/events/aie-shanghai-2026/s07-stepfun-agent-harness/Thu, 05 Nov 2026 13:45:00 +0800https://yousali.com/events/aie-shanghai-2026/s07-stepfun-agent-harness/长程 agent 能否交付可信结果，模型只是底座；还要靠 harness 补上保存进度、挑选上下文、事先定验收条件、云端持续执行这几个工程环节。圆桌：Community in the AI erahttps://yousali.com/events/aie-shanghai-2026/s06-panel-community-ai-era/Thu, 05 Nov 2026 13:00:00 +0800https://yousali.com/events/aie-shanghai-2026/s06-panel-community-ai-era/AI 时代信息差和产品差异变小以后，线下见面带来的信任和人与人的连接反而更值钱；创业者对外发声要像做产品一样准备，攒够了再说。Measure the frontier of AI for real world usehttps://yousali.com/events/aie-shanghai-2026/s05-arena-measure-the-frontier/Thu, 05 Nov 2026 11:25:00 +0800https://yousali.com/events/aie-shanghai-2026/s05-arena-measure-the-frontier/Arena 自己做通用 agent 产品吸引真实使用，每个会话随机分配底层模型，从真实轨迹里挖信号合成排行榜；按这个榜，美国前沿实验室领先，中国开源权重模型在第二梯队但处在性价比前沿。Full-Stack Inference Engineering: Latency, Throughput and Costhttps://yousali.com/events/aie-shanghai-2026/s04-baseten-full-stack-inference/Thu, 05 Nov 2026 11:10:00 +0800https://yousali.com/events/aie-shanghai-2026/s04-baseten-full-stack-inference/做好开放模型推理要先定延迟、吞吐、成本的取舍，再把单 GPU 引擎优化和跨云集群调度叠起来，每一层都必要但单靠哪一层都不够。Post-Training at the Frontier: Teaching Open Models to Search and Browse the Webhttps://yousali.com/events/aie-shanghai-2026/s03-youcom-post-training-search/Thu, 05 Nov 2026 10:50:00 +0800https://yousali.com/events/aie-shanghai-2026/s03-youcom-post-training-search/用中等热度网页自动出题、按难度筛过的小规模 RL 数据集，就让一个开源 30B 模型在 BrowseComp 上明显变强，效率也作为副产品变好。When we can do anything, do we still know what we want?https://yousali.com/events/aie-shanghai-2026/s02-ryo-lu-what-we-want/Thu, 05 Nov 2026 10:25:00 +0800https://yousali.com/events/aie-shanghai-2026/s02-ryo-lu-what-we-want/工具越强，人越容易把答案完整当成自己想过；好工具应该像工作台，让人碰得到正在改的东西，而知道自己想要什么是在动手改的过程中长出来的。Building Agents: Yesterday, Today, Tomorrowhttps://yousali.com/events/aie-shanghai-2026/s01-langfuse-building-agents/Thu, 05 Nov 2026 10:10:00 +0800https://yousali.com/events/aie-shanghai-2026/s01-langfuse-building-agents/agent 架构正在收敛到 coding agent 的样子，trace 多到人读不过来，改进循环要交给后台 agent 跑，人只保留对什么算好、发不发的判断。From Models to Working Enterprise Agentshttps://yousali.com/events/aie-shanghai-2026/m02-microsoft-enterprise-agents/Thu, 05 Nov 2026 09:45:00 +0800https://yousali.com/events/aie-shanghai-2026/m02-microsoft-enterprise-agents/企业 agent 接下来的关键变化是拿到自己的身份、按最小权限单独授权；agent 之间会不经编排地互相调用，测试要转向 rubric 专项评测和模拟世界隔离测试。Opening with swyxhttps://yousali.com/events/aie-shanghai-2026/m01-swyx-opening/Thu, 05 Nov 2026 09:35:00 +0800https://yousali.com/events/aie-shanghai-2026/m01-swyx-opening/主持人欢迎和赞助致谢，加上 swyx 的视频问候：AIE 面向真正做东西的工程师，这是第一次来中国，也是第一场不直播的 AIE。