智东西（公众号：zhidxcom）
作者 | 江宇
编辑 | 李水青
智东西10月29日报道，今日，在2026腾讯全球数字生态大会上，腾讯集团高级执行副总裁、云与智慧产业事业群CEO汤道生发表主旨演讲，围绕AI Agent在企业的规模化落地、腾讯云AI产品与基础设施升级，以及全球化战略等话题，分享了腾讯AI Agent及基础设施建设的最新进展。
汤道生宣布，腾讯混元大模型Hy4正式版即将发布。相比今年8月28日发布并开源的Hy4 Preview，新版本实现全面提升，尤其强化了长程任务和真实生产力任务的处理能力，显著提升思维链效率，并首次加入视觉理解能力。腾讯还将于近期推出Hy4 Flash，在继承Hy4视觉理解、Agent等能力的基础上，进一步追求极致性价比。
在AI应用方面，腾讯正式发布跨应用AI协同体系Buddy One及腾讯网盘。
Buddy One将打通腾讯文档、腾讯会议、腾讯网盘、ima等办公协同产品，共享WorkBuddy Agent底座，并统一账号和积分池。后续还有更多产品加入Buddy One。
腾讯网盘则是一款面向Agent时代的全新记忆空间产品，为WorkBuddy套件提供统一存储，可以让Agent在工作中积累的上下文，以及生成的成果，安全保存、随时调用，沉淀为组织资产。当前，WorkBuddy的资料库能力，就是由腾讯网盘提供的支撑。
同时，腾讯云将进一步打通桌面办公智能体WorkBuddy、企业微信与企业内部管理系统，让Agent从辅助生成内容，走向参与业务分析、经营决策和任务执行。目前，WorkBuddy已累计迭代100个大版本，完成90亿次用户任务请求。
谈及企业AI的发展方向，汤道生认为，企业使用AI的关键，正在从“用上”转向“用好”。企业需要的不仅是更强的模型，还需要能够深入业务流程、持续创造生产力、提升经营效率和投资回报的Agent系统。
此外，汤道生披露，过去一年，腾讯云国际业务海外活跃客户数量增长约3倍，海外AI需求实现三位数增长。腾讯云将加快全球数据中心布局，并推动WorkBuddy、ADP、Agent Runtime等AI产品出海。
此次演讲的核心信息与金句提炼如下：
1、Agent爆发：今年是Agent使用全面爆发的一年，员工使用AI Agents提效，已经是必选题。
2、组织变革：腾讯内部已实现Agent办公覆盖全员、全场景。腾讯正在培养和识别“超级个体”的同时，探索智能驱动的新组织方式。
3、经营决策：在AI时代，速度本身就是竞争力。
4、模型与成本：企业AI应用正在从粗放使用最强、最贵的模型，转向根据任务精准匹配不同模型，追求更高性价比。腾讯混元Hy4正式版及Hy4 Flash即将发布。
5、全球化：中国科技在全球的竞争力，已经不只是低成本，更多来自AI带来的效率，以及云与工程的综合交付能力。腾讯云希望成为连接中国AI与全球产业的桥梁。
汤道生演讲实录如下，智东西做了不改变原意的编辑：
各位嘉宾上午好，很高兴和大家再次相聚腾讯全球数字生态大会。
过去两年，AI的发展超出了几乎所有人的预期，今年更是Agent使用全面爆发的一年。国家数据局统计，我国日均Token调用量，已经从2024年初的1000亿，增长到今年7月的超500万亿，两年增长超5000倍。
这一数据的背后，是AI应用的快速发展，Agent正在成为新的生产力形态。
以WorkBuddy为代表的AI办公智能体，除了日常文档协作外，在商业分析、业务运营、市场调研、营销规划与执行等专业领域，正在各行各业快速落地。员工使用AI Agents提效，已经是必选题。企业也需要AI做更准的决策与更快的执行，加强在市场上的竞争力，否则不进则退。
从今年3月上线至今，WorkBuddy的Token调用量持续攀升，企业版落地50多个行业，覆盖中国企业500强中的50%、企业100强中的60%。开放平台的伙伴和开发者数量也超过5万家，上架技能（Skills）超过20万。同时，我们看到，WorkBuddy在用户留存、使用时长、长程任务处理等关键指标上，也都有不错的表现。
我自己就是元宝与WorkBuddy的重度用户。准备今天的演讲时，我就把思路喂给WorkBuddy，它帮我整理资料、提炼框架、生成建议，我们一边讨论一边修改，实现了“人机共创”。
在腾讯内部，用Agent办公也已经覆盖全员和全场景。WorkBuddy深入了多个部门，连接了内部系统，同事们每天用它完成工作，生成与分享报告，并沉淀了不少大模型可持续调用的专业技能，当然也给WorkBuddy产品团队提供了很多反馈与优化建议。我们定期组织WorkBuddy使用案例的跨部门分享，举办内部黑客松，鼓励大家探索出更多有价值的AI应用。腾讯云在培养与识别超级个体同时，也在探索新的组织方式。
在打造“智能驱动新组织”的过程中，腾讯云不断完善智能体产品矩阵，既满足不同业务场景的需求，也提供低成本、高性能、安全可控的多Agent运行环境，打造“全场景、高效能、更放心”的Agent生产力系统。
接下来，我就围绕这三个关键词，具体分享下腾讯的探索和实践。
一、打通办公工具与业务系统，让Agent理解业务上下文、融入工作流程
第一个词是“全场景”。
企业客户经常反馈：模型已经很强，工具也不少，但真正用到业务里，仍然有很多的断点。比如，AI能做分析、写方案，如果接不上业务系统拿数据，就无法了解企业的业务、渠道和客户的实际情况。
要打通这些断点，首先要做好连接。WorkBuddy就是把人、工具、系统和数据连接起来，让大模型理解业务上下文，融入实际工作流程。
1、发布Buddy One与腾讯网盘，统一调用办公工具、保存工作成果
首先是打通办公场景，让工具围绕任务协同起来。
过去几十年，数字化办公主要以文件为载体，不同类型的文件格式，由不同应用来浏览与编辑，都是为人们使用鼠标与键盘输入来设计的。桌面工具多，文档太零散，编辑时间比较长。在多人协作时，各自修改自己的版本、再来回传阅，效率低，也挺麻烦。
但在办公Agent时代，我认为多人协作将会进入人机共创的新范式：团队讨论完把需求提交给AI，内容由大模型生成，由Agent来整合，团队来验收，然后沉淀到知识库。
今天，借生态大会的机会，我们也正式对外发布Buddy One。
Buddy One以WorkBuddy为工作台，无缝整合腾讯文档、腾讯乐享、腾讯设计等不同工具，通过统一账号、统一积分池、统一数据池，大模型针对不同需求，调用不同工具来生成多模态内容。团队可以在项目中协同修改，共同检验产出物，持续讨论与迭代。Buddy One实现人和AI共同创作，围绕任务协同，而不是围绕文件流转。当然，还要支持好多终端的联动，云与端的Agent无缝切换。
这样的“一体化体验”，需要多个产品共同支撑。腾讯文档负责内容编辑，腾讯乐享提供知识管理，腾讯会议把开会的共识融入代办中，皆为WorkBuddy提供了上下文。
当任务需要补充外部信息时，元宝搜索也会提供实时数据。依托国内外的权威网站、专业的数据资源、腾讯独有的内容生态，如腾讯医典等，元宝搜索沉淀了千亿级的索引，满足各类深度调研的需求。元宝既作为ChatBot给用户提供快而准的答案，也通过API向行业提供搜索能力。
今天，我们也将正式发布“腾讯网盘”，这是一款面向Agent时代的全新记忆空间产品，为WorkBuddy套件提供统一存储，可以让Agent在工作中积累的上下文，以及生成的成果，安全保存、随时调用，沉淀为组织资产。当前，WorkBuddy的资料库能力，就是由腾讯网盘提供的支撑。
2、WorkBuddy接入企业微信与经营系统，支持智能问数
其次是打通数据平台，让Agent进入企业工作流。
过去，企业的各种管理系统如ERP、CRM，沉淀了大量企业经营的数据，有特定UI界面来查看数据与管理流程，但每个系统割接，虽然各自有些API接口，但跨系统的数据分析，就算系统打通了，工作量巨大，每次等待时间长。
今天使用WorkBuddy，通过让这些管理系统开放MCP接口，只要建立好对应的skills，描述清楚数据结构与关系，Agent就能拆解各种业务分析的任务，主动向多个系统获取需要的数据，自动生成所需要的报表，呈现分析结果。当然，我们也可以搭建不同任务Agent，异步来向系统提交数据，比如批量把发票整理好，批量提交报销，不需要像以前那样，同步操作UI界面，又慢又繁琐。
还有很多企业，员工在企业微信讨论项目分工，达成的共识，这对办公Agent都是很有重要的上下文输入。
就在这个月，WorkBuddy和企业微信联合发布了新版本，实现了AI办公与移动办公的融合与升级，WorkBuddy正式以智能机器人身份入驻企业微信侧边栏，员工可以像跟同事对话一样，随时把任务交给它。WorkBuddy可以24小时开工，做完把结果直接发回企微。员工也可以用企微账号登录WorkBuddy，授权它读取数据、执行任务。
就像DEMO演示的，通过WorkBuddy APP发出一个简单指令，它就能在授权范围内，查看未读消息和邮件，整理出重点事项和待办，推送到手机上；需要时，还能生成回复建议，经过确认后一键发送。对于高频出差、会议密集的员工来说，用碎片时间就能高效处理工作。
WorkBuddy与企微打通后，用户在企微里的沟通、文档和会议记录，也都能在授权范围内，成为Agent理解任务、开展工作的依据。企业不用换系统、搬数据、重做流程，直接可以把沉淀多年的数字化资产，变成AI的生产资料。管理者还可以统一分配使用名额、Token额度和访问权限，员工离职后统一收回。
同时，WorkBuddy也会继续开放产品架构，支持企业自研Agent与生态中第三方Agent，包括第三方IM系统的接入。选什么模型、接什么Agent、用什么IM，选择权都在企业手里。越来越多企业正在基于WorkBuddy组成自己的企业Agent网络，搭建新一代的企业智能中枢。
刚才也提到，Agent要掌握更多企业数据与上下文，还必须进一步连接企业的内部管理系统，比如ERP、CRM、OA和财务软件等等。
以腾讯云为例，我们打造了经营服务智能体“磐石小O”。过去，我们同事要想查客户的经营数据，需要找系统、走审批、翻报表；现在，在WorkBuddy里问一句小O，就能在权限范围内，查到销售业绩、客户经营情况和回款进度。小O依托WorkBuddy的“托管智能体服务”（WMA），通过统一身份认证、权限管控和7×24小时隔离环境运行，让Agent的每一次操作都可审计、可追溯，确保经营数据安全。
智能问数，对于管理决策非常有价值。我自己深有体会。比如过去，数据工程师做一张报表要两三周，管理者看到后，想追问业务细节、寻找优化方向，又得再等报表更新，很容易错过决策时机，如果仓促决定，决策质量就很难保证。现在，AI原生的管理者可以通过Agent随时提问，及时了解业务动态，更快看清问题、作出判断、采取行动。对管理者来说，这种决策提效，是企业保持长期竞争力的关键。谁能更好使用AI作出更快决策、更早行动，就可能在市场更早发现商机，或者在存量市场获得更多的份额。在AI时代，速度本身就是竞争力。
对于每个员工来说，Agent打通经营系统后，不只是能查数据，还能办更多的事情。比如销售在拜访客户前，小O可以根据CRM数据和公开信息，整理客户动态和沟通建议；一句话预约会议、同步日程；拜访后，再汇总业务数据，生成报告和待办。需要审批时，也能直接发起。
3、MAGIC串联营销与下单，ADP支持定制服务Agent
最后是构建业务Agent，执行运营任务，连接伙伴，服务客户。
从内部协同，到业务经营，Agent也可以针对清晰目标与可量化指标，持续执行任务。比如在客服、营销、交易等场景，企业需要专业的业务Agent：它们遵循业务规则与流程、全天候运行，不只完成任务，更要服务于经营者，对实际的经营结果负责。
最近，腾讯企点营销云MAGIC智能体已在WorkBuddy开放平台上线Buddy应用。以热门商品促销为例，WorkBuddy整理商品清单，MAGIC连接客户数据平台CDP，圈选人群并执行营销活动，云Mall承接下单，一个入口即可串起准备、执行、成交和复盘全流程。
企业也可以通过腾讯云ADP，定制更加个性化的服务Agent，发布到企业智能体统一门户AgentPortal，或者直接嵌入到企业官网、小程序和业务系统。我们也在进一步推动ADP与WorkBuddy的深度打通，让更多业务Agent接WorkBuddy，成为员工随手可调用的能力。
在伊利，他们正在组合使用ADP、WorkBuddy和企业微信，做营销和销售：WorkBuddy汇聚业务数据，分析并生成经营策略，下发给上千名导购，运营提效超过50%；ADP导购智能体则可以实时生成商品推荐和沟通话术，A/B测试显示，销售订单提升26%。
二、混元Hy4将加入视觉理解，Flash主打性价比，Agent Runtime同步升级
接下来，我们讲第二个词：“高效能”。
随着Agent持续渗透办公与业务场景，并开始支撑企业级的生产，企业的Tokens使用量与成本暴增，迫切需要更加精细化的tokens资源管理，降低成本。从粗放使用最强、最贵的模型，转向追求性价比，精准匹配不同模型，来执行不同任务。
腾讯云拥有高性能与极高性价比的Agent基础设施，稳定的异构算力集群，与多家芯片厂商深度合作，构建国产算力的新生态。
在全球算力紧张情况下，腾讯云深度适配多家厂商的芯片，持续优化国产芯片上的推理性能，达到很好的性价比，在具身智能等制造行业得到广泛的使用。为承接Hy4 Preview发布后的调用洪峰，腾讯云与混元团队紧密合作，在上线前两周，把某国产芯片单卡吞吐量提升了5倍；我们还帮助汽车制造国际头部Tier1，落地国产千卡集群，支撑端到端智驾大模型的规模化训练。
模型上，腾讯混元Hy持续在真实业务场景中进化。过去一年，Hy重建了数据和训练基础设施，平均两个月发布一个大版本。Hy4 preview定位成“为生产力而生”，在代码、办公、科学等真实生产力任务上，非常稳定与靠谱，在发布后一直在OpenRouter排行前列，支撑了元宝、WorkBuddy等产品的稳定服务。
混元大模型Hy4很快会发布，相比Hy4 preview实现了全面大幅提升，尤其强化了各类长程和真实生产力任务的能力，显著提升了思维链效率，并首次引入了视觉理解。
Hy4 flash也会在近期发布，提供极致性价比的实用模型，在继承Hy4视觉、Agent等强大能力的基础上，通过模型和产品的深度Co-Design高频进化。
腾讯始终坚持开放的策略，基于TokenHub平台，用户可通过API调用混元、DeepSeek、GLM、MiniMax、Kimi等主流开源大模型。配合Token Plan统一计费服务，企业能以极低切换成本，在多模型之间按需选用。
在内容生产场景，我们也推出腾讯云音视频AI WAND产品家族，提供主流的多模态模型、WAND-Vega多模态应用模型，与100余款自研的媒体处理工具，覆盖了90%的头部短漫剧出海客户，日均生成450万张图片、300万秒视频。
另外，在Agent Infra方面，今天我们也全新升级了Agent Runtime，围绕上下文管理、工具调用、安全沙箱、可观测、评测等关键环节做了系统性增强，让Agent跑得更稳、更可控，综合性价比更高，稍后我的同事会做详细介绍。
三、拦截高危操作、记录调用过程，企业数据支持加密与恢复
最后一个关键词是“更放心”。
随着越来越多Agent在办公网内部署，Agent与大模型都会触碰到更多企业级数据，保障网络、运行环境与数据的安全，将变得尤为重要。企业不仅要防止敏感数据泄漏给第三方，也要防止内外部用户的恶意攻击，还有知识产权保护、超越权限范围的数据窃取。
AI把坏人的能力同样放大，Agent把作恶的门槛也大大降低，最近我们看到企业根密钥被窃取的数量暴增，后来找到我们安全团队修复问题，并部署新一代安全产品加强防护。
腾讯自身经营着许多国民级的基础服务，我们知道安全是信任的基础，因此始终把安全放在产品和服务首要考量。面向Agent场景，我们致力于构建“可信AI”治理体系，主要包括三个方面。
一是行为可控。Agent代表谁、能看什么、能做什么，先定义清楚；运行中能监控、能干预、能叫停。在 WorkBuddy中，隐藏的恶意指令会被识别清洗，扩展能力加载前要通过安全校验，删除数据、越权访问等高危操作直接拦截，不确定的操作交给用户确认。
二是过程可溯。每个Agent都有独立身份，纳入企业统一账号和权限体系；看过什么数据、调用过什么工具、执行过什么动作，全程留痕，管理员按人、按时间、按Agent就能定位问题，也可以统一配置规则，限制数据外发。
三是数据可管。企业最不放心的是数据：存在哪、怎么传、误删了怎么办？我们的答案是全程加密、兜底恢复：私有数据从终端、传输到存储全程加密，敏感任务支持机密计算；万一误删、覆盖或损毁，可做文件级恢复，不让一次意外变成不可挽回的损失。
这些能力也得到了外界的广泛认可。WorkBuddy是国内首批通过信通院“可信AI”评估的产品，也获得了CCIA（中国网络安全产业联盟）“个人信息保护最高级评价”。同时，我们还作为核心单位，参编了深圳《人工智能可信办公智能体标准体系》，为标准制定提供了可信AI办公最佳实践。这些安全能力建设，也赢了广东省政数局、深圳航空、深圳通等政企客户的认可。
我们始终认为，安全不让位于效率，创新才能走得更远。腾讯将持续把产品做好，也把信任做深，成为企业可信赖、可托付的AI伙伴。
四、海外AI需求三位数增长，加快数据中心布局与Agent产品出海
好，讲完腾讯AI的全场景、高效能、更放心。接下来，我想跟大家分享一下腾讯云在全球化服务的进展与对未来的思考。
腾讯也在积极把国内大模型与AI Agent的落地经验带到全球，从算力基座、模型服务，到Agent驱动的技术平台与行业解决方案，帮助更多全球企业搭建高性价比、安全可控的AI技术栈。
过去一年，腾讯云国际业务海外活跃客户数翻倍，覆盖了全球超150个国家和地区。在几个重点市场，我们都跑出了业务纵深：
在印尼，我们先后签下并积极推进XLSMART、Telkomsel、Indosat三大本地运营商，在云与AI的全面合作。
在泰国，我们持续为正大集团更多业务板块降本增效，加速智能化转型，并进入头部运营商、交易所和银行。
在港澳，我们服务了鸿福堂、政府数字政策办公室等客户，业务连续七年高位稳健增长。
在欧洲，我们持续扩展与iyzico、eMAG等全球互联网平台企业的深度合作。
面向未来，AI正在重塑全球云市场的选择逻辑，客户选云，不再只看资源、成本和稳定性，更看模型能力和AI工具链。
中国AI在产业场景、产品体验和性价比上具备领先优势。腾讯云正在把这些能力，带向全球，既助力中企出海，也服务全球客户。
首先，我们会加快全球数据中心建设。
腾讯云海外开区速度长期保持行业领先，全球已部署超过200万台服务器。过去一年，我们在法兰克福、马来西亚、墨西哥、大阪、利雅得和泰国，新增或扩容了数据中心。未来一年，我们还将进一步拓展亚洲、中东、欧洲和美洲的布局。
其次，是打造高性价比的基础设施。
面对着AI大浪潮带来数据，通算和存储需求的激增，我们持续为客户设计最优的软硬件架构。比如星星海最新旗舰整机，支持1024核，云支撑云服务器升级到最新的SA10,加上极致的内存管理技术，为客户提供极具性价比的算力平台。
面对存储成本的上涨，我们也为许多客户设计分层存储方案；当前COS磁带存储累计超2000PB，帮助全球客户低成本归档“极冷数据”。
对于拥有自己数据中心，并要从虚拟化技术升级到云原生基座的客户，专有云TCE最小部署规模，从14台降到8台，进一步降低部署门槛与成本。
此外，我们也敏锐把握AI应用机遇，提升全球交付能力。
过去一年，腾讯云的海外AI需求实现了三位数增长，而且全球企业对我国的AI产品生态有着极高的兴趣与需求。
模型层面，国产开源模型在办公、知识问答、内容生成等高频场景，展现出很好的体验与极高的综合性价比，Hy4 Preview上线后，很快收到众多海外企业的咨询。未来，我们也将投入更多资源，保障优质Token的全球供给。
产品层面，我们加速推动CodeBuddy、WorkBuddy、腾讯设计Miora、ADP、MAGIC AI、Agent Runtime等各类Agent全线出海，无论中国企业的外派员工，还是海外客户的本地团队，都能在同一套AI工作台上协同作业。
服务海外客户方面，我们持续协助海外平台企业打造本地Superapp生态，升级版的小程序平台方案Superapp as a Service与CodeBuddy实现了深度协同，开发者用自然语言，就能生成小程序。印尼本土运营商Telkomsel基于它，构建小程序生态，MAU一年内翻了一番。
在音视频场景，今年美加墨世界杯期间，WAND音视频AI技术，为17个国家和地区的官方转播平台提供了技术服务。名古屋亚运会期间，国际奥委会主席现场点赞了我们和央视总台共创的中国转播技术。
在云迁移场景，我们基于CodeBuddy和WorkBuddy开发了20多个迁移Skills，只用4.5个月，就帮助印尼头部运营商XLSMART完成平稳迁移，同时还开发了200多项本地化功能。
今天，中国科技企业的全球竞争力，已经不只来自成本，更来自AI、云和工程交付的综合能力。腾讯云希望成为连接中国AI与全球产业的桥梁，帮助更多企业在全球市场获得增长。
各位嘉宾、各位伙伴，回望信息技术的发展，真正的产业变革，不只是来自技术突破，更来自技术深入业务，加强与上下游连接，甚至改变组织方式。
AI也一样。从“用上”到“用好”，关键是让AI进入流程、创造生产力，持续提升经营效率与投资回报。腾讯将坚持长期投入，持续打磨产品、开放平台能力，助力企业打造可运营、可进化的Agent生产力系统。
这个过程中，我们也始终向伙伴敞开大门，期待与大家一起，共同打造繁荣的Agent产业生态。
谢谢大家。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
Author: Zheng Zheng, Jiang Woo
Editor, Lee Shui-qing
On October 29th, Shigeki reported that today, at the 2026Tencent Global Digital Ecology Congress, the Senior Executive Vice-President of the Tencent Group, CEO of the Clouds and Wisdom Enterprise CEO, delivered a keynote address to share the latest developments in the development of Tencent AI Agent and infrastructure around the issues of AI Agent ' s s scalding of enterprises, Tencent cloudai products and infrastructure upgrading, and globalization strategies.
Thomason announced that the official version of Tencent hybrid LLMHy4 would be publish. Compared to the publish and open-source Hy4 Preview of 28 August this year, the new version has been comprehensively upgraded, in particular by enhancing the processing capacity of long-range and real-productivity missions, significantly increasing the efficiency of the thinking chain and, for the first time, incorporating visual understanding. Tencent will also be introducing Hy4 Flash in the near future to further pursue value for money based on the ability to inherit Hy4 visual understanding, Agent, etc.
In terms of AI applications, Tencent formally publish cross-applies the AI Synergic System Buddy One and Tencent Web.
Buddy One will access Tencent files, Tencent meetings, Tencent webdisk, and Ima, among other office co-products, share the WorldBuddy Agent base and harmonize the account numbers and scorecards. More products were added to Buddy One.
The Tencent web drive is a completely new memory space product for the Age of Agent, providing a unified storage of the WorkBuddy package, allowing Agent to accumulate context in its work, as well as the results generated, to be safely preserved, called at any time, and deposited as an organizational asset. Currently, the WorkBuddy database capacity is supported by Tencent grid.
At the same time, Tencent clouds will provide further access to desktop office AI agentWorkBudy, enterprise WeChat and enterprise internal management systems, allowing Agent to move from supporting content generation to participation in business analysis, business decision-making and task execution. Currently, WorkBuddy has cumulatively completed 100 large versions, with 9 billion user job requests.
Turning to the direction of corporate AI, Thomason believes that the key to the use of AI by enterprises is shifting from “use on” to “use on”. Enterprises need not only stronger models, but also Agent systems capable of deepening business processes, generating sustained productivity, improving business efficiency and investment returns.
In addition, Thomason has revealed that the number of active overseas clients in Tencent Clouds International has increased by about three times over the past year, with a three-digit increase in offshore AI demand. Tencent clouds will speed up the global data centre layout and facilitate the shipping of a number of AI products such as WorldBuddy, ADP, Agent Runtme.
The core message and sentence of the presentation are as follows:
1Agent outbreak: This year is the full-blown year of Agent's use, and the use of AI Ages by employees is a mandatory theme.
2. Organizational change: inside Tencent, Agent ' s office coverage has been achieved. Tencent is exploring new forms of organization driven by intelligence while developing and identifying “super individuals”.
Business decision-making: In the AI era, speed itself is competitive.
Models and costs: Enterprise AI applications are moving away from using the strongest and most expensive models in rough print to matching different models with mission precision in pursuit of higher value for money. Tencent mixed Hy4 and Hy4 Flash will be publish.
5. Globalization: China ' s global competitiveness in science and technology has become not only low cost, but more of the efficiency brought about by AI, as well as the combined ability to deliver clouds and engineering. Tencent clouds want to be a bridge between China's AI and global industry.
The following is a video of the talk by Tom Doo-sheng, whose intellectuals have been edited without changing their intentions:
Good morning, and it is a pleasure to meet you again at the Tencent Global Digital Ecology Conference.
Over the past two years, AI has developed beyond almost everyone's expectations, and this year has been a full-scale outbreak for Agent. According to the National Data Agency, the daily average number of token calls in the country has increased from 100 billion at the beginning of 2024 to over 50 trillion in July this year, more than 5,000 times the increase in two years.
Behind this data is the rapid development of AI applications, and Agent is becoming a new form of productivity.
In addition to day-to-day documentation collaboration, the AI office AI agent, represented by WorkBuddy, is rapidly landing in a variety of fields of expertise, including business analysis, business operations, market research, marketing planning and implementation. The use of AI Agents by employees is already a mandatory subject. Enterprises also need better decision-making and faster implementation to enhance competitiveness in the market, or fail to advance.
From March this year onwards, Token calls in WorkBuddy have continued to climb, with more than 50 industries on the ground, covering 50 per cent of the top 500 Chinese enterprises and 60 per cent of the top 100 enterprises. There are also more than 50,000 partners and developers of open platforms, and over 200,000 Skills. At the same time, we see that WorkBuddy also performed well on key indicators such as user retention, length of use, and long-range task processing.
I am a serious user of Yuanbao and WorldBuddy. When I was preparing for today's presentation, I gave my thoughts to WorkBuddy, which helped me to sort out the information, refine the framework and generate proposals, and we changed it while we were discussing it, and we realized the idea of “human creation”.
Inside Tencent, the office in Agent has also covered the whole staff and the whole scene. WorkBuddy has in-depthd many departments, connected internal systems, used daily by colleagues to carry out their work, generate and share reports, and sunk a number of LLM professional skills for sustainable calls, and, of course, provided a lot of feedback and optimization to the WorkBuddy product team. We regularly organize cross-sectoral sharing of cases of WorldBuddy use and internal hacking to encourage more valuable AI applications to be explored. Tencent clouds are exploring new ways of organizing while nurturing and identifying super individuals.
In the process of building an “intellectually driven new organization”, the Tencent cloud continues to refine the AI agent product matrix, both to meet the needs of different business scenarios and to provide a low-cost, high-performance, safe and manageable multi-Agent operating environment and to create a “wide-scape, efficient, and more secure” Agent productivity system.
Next, I will share the exploration and practice of Tencent in the context of these three key words.
i. Access to office tools and business systems to enable Agent to understand operations context and integrate workflow Cheng
The first word is “the whole scene”.
Regular feedback from business clients: Models are already strong and tools are available, but there are still many break points that are actually used in business. For example, AI is able to analyse and write programmes, and if it does not have access to data in the business system, it does not have access to the actual business, channels and clients of the enterprise.
To get through these breakpoints, you need to get connected first. WorkBuddy connects people, tools, systems and data, allowing LLM to understand business context and integrates it into the actual workflow process.
1, publishBuddy One and Tencent Web Disk to harmonize office tools and save work
The first is to create an office landscape where tools are aligned around mandates.
Over the past few decades, the digital office has been mainly a file carrier, different types of file formats, browsing and editing by different applications, designed for people using mouse and keyboard input. There are many desktop tools, documents are too fragmented, and editing time is longer. It is also inefficient and cumbersome for each to modify its own version and re-transmit it back when many people work together.
But in the age of Agent, I think that multi-person collaboration will enter a new paradigm created by human beings: After discussion, the team submitted the demand to AI, which was generated by LLM, integrated by Agent, the team came to acceptance check and then settled into the knowledge base.
Today, on the occasion of the Eco-Congress, we are also officially outside publishBuddy One.
Buddy One, using the WorkBudy as a desk, seamlessly integrates different tools, such as Tencent documents, Tencent pleasures, Tencent designs, which generate multi-modular content through unified account numbers, unified scorecards, unified data pools, LLM for different needs, and calls on different tools. Teams can modify the project in concert, test the output together, and discuss and overlap on an ongoing basis. Buddy One brings people and AI together to work together around tasks rather than documents. And, of course, there's a lot of terminal connections that have to be supported, and the cloud has to be seamlessly switched to the edge of Agent.
Such an “integrated experience” needs to be supported by multiple products. Tencent documents are responsible for content editing, Tencent pleasures provide knowledge management, and Tencent meetings incorporate consensus of the meeting into the Chargé d ' affaires, providing context to WorkBuddy.
When a mission requires additional external information, a betel search also provides real-time data. Based on authoritative websites at home and abroad, specialized data resources, and the unique content ecology of Tencent, such as the Tencent Medical Code, the Treasure Search has deposited hundreds of billions of scales of indexing to meet the needs of various types of in-depth research. Yuanbao provides both quick and accurate answers to users in ChatBot and industry search capabilities through API.
Today, we will also have the official publish “Tencent Netboard”, an entirely new memory space product for the Age of Agent, which will provide a unified storage of the WorldBudy package, allowing Agent to accumulate context in its work, as well as the results generated, to be safely preserved, called and deposited into organizational assets. Currently, the WorkBuddy database capacity is supported by Tencent grid.
2. WorkBuddy access to enterprise WeChat and operating systems to support smart questions
The second was to access the data platform and get Agent into the enterprise workflow.
In the past, enterprise management systems such as ERP, CRM have deposited a large amount of data on business operations, with specific UI interfaces to view data and management processes, but each system has been cut down, although there are some API interfaces, but data analysis across the system, even if the system is accessible, has a heavy workload, with long waiting times.
Using WorkBuddy today, using these management systems to open up the MCP interface, Agent can dismantle the tasks of various business analyses by creating a corresponding skills, describing the structure and relationships of the data, proactively obtaining the required data from multiple systems, automatically generating the required reports and presenting the results of the analysis. Of course, we can also create different tasks, Agent, to move forward, to submit data to the system, for example, through bulk preparation of invoices, bulk submission of claims, without having to operate the UI interface in parallel, as was the case before.
There are also a number of enterprises where employees discuss the division of work and a consensus has been reached on the division of work of the project, which is an important context entry for Agent.
Just this month, a new version of the joint publish version of WorkBuddy and the WeChat enterprise was created to integrate and upgrade AI and mobile offices, and WorkBuddy formally entered the sidebar of the WeChat company as a smart robot, where employees can give their assignments as they speak to their colleagues at any time. Work Buddy can start work 24 hours and send the results back directly to the company. The employee can also log in with his or her micro-accounts to authorize him or her to read the data and perform his or her tasks.
As demonstrated by DEMO, a simple instruction is given through the WorldBudy APP, which allows it to view unread messages and e-mails, organize priorities and to-dos, and push them to a mobile phone, within the limits of its mandate; it also generates suggestions for responses, which are sent after confirmation. For staff with high frequency travel and intensive meetings, debris time allows for efficient disposal.
When WorkBuddy communicates with the company, the user ' s communication, documentation and minutes in the enterprise can also, within the mandate, form the basis of Agent ' s understanding of the mission and work. Instead of changing systems, moving data and re-engineering processes, enterprises can directly transform digital assets that have been deposited for many years into AI production data. The managers could also make a uniform allocation of use places, Tokenquota and visit permission, which would be recovered by staff members after separation.
At the same time, WorkBuddy will continue to open up product structures to support enterprise self-study of access to Agent and ecological third-party Agent, including third-party IIM systems. The choice is in the hands of the firm. An increasing number of enterprises are forming their own business Agent network based on the WorldBudy, building a new generation of enterprise intelligence hubs.
As was mentioned earlier, in order for Agent to have more business data and context, it is also necessary to further connect the enterprise ' s internal management systems, such as ERP, CRM, OA and financial software.
In the case of the Tencent cloud, we built the AI agent Rock O. In the past, our colleagues had to look for systems, go through approvals and flip statements to check their clients' business data; now, asking O-minus in the WorldBuddy, we can find sales performance, customer operations and the progress of repayments within permission. The “Hotting AI agent Service” (WMA) of Little OitoWorkBuddy, which operates in a single identity, permission control and 7x24-hour segregated environment, allows every operation by Agent to be audited, traceable and secure for operating data.
Smart questions are very valuable for management decision-making. I knew it myself. In the past, for example, a data engineer had to make a statement for two or three weeks, and when the manager saw it, he wanted to ask for business details, look for an optimal direction and wait for the report to be updated, he could easily miss the timing of the decision-making process and, if he had taken a hasty decision, the quality of the decision would be difficult to ensure. It is now possible for AI original managers to ask questions at any time through Agent, to keep abreast of business developments and to see, judge and act more quickly. For managers, the effectiveness of such decisions is key to the long-term competitiveness of enterprises. Those who are better able to use AI for faster decision-making and earlier action may find business opportunities earlier in the market or gain a greater share in the stock market. In the AI era, speed itself was competitive.
For each employee, Agent has access to the business system, not just data, but more. For example, before a visit to a customer, Junior O can organize client dynamics and communication suggestions based on CRM data and public information; book a meeting in one sentence, synchronize the calendar; and, after a visit, aggregate business data, generate reports and to-dos. When approval is required, it can also be initiated directly.
3. MAGIC Serial Marketing and Listing, ADP supports customized services for Agent
Finally, the construction of the business Agent to perform operational tasks, connect partners and serve clients.
From internal synergy to business operations, Agent can also carry out its tasks on an ongoing basis with clear objectives and indicators that can be quantization. For example, in the context of hospitality, marketing and trading, enterprises need professional business Agent: they follow business rules and processes, operate 24 hours a day, not only to complete their tasks, but also to serve the operator and be responsible for the actual results of their operations.
Recently, the Tencent enterprise marketing cloud MAGICAI agent has been put on site on the WorkBuddy Open Platform. In the case of hot commodity promotions, for example, the WorldBuddy catalogues, MAGIC links the client data platform CDP, crowds and marketing activities, cloud Mall takes over the list, and a single entry point allows for a full process of preparation, execution, transaction and round-up.
Through Tencent CloudADP, enterprises can also customize a more personalized service, Agent, publish, to the Enterprise AI agent Consolidated Portal, Agent Portal, or be embedded directly into the corporate network, small programs and business systems. We are also further promoting the outreach of ADP to WorldBuddy, making it possible for Agent to connect to the WorldBuddy and become a ready-to-call capability for staff.
In Erie, they are combining ADP, WorkBuddy and Enterprise WeChat for marketing and marketing: WorkBuddy collects business data, analyses and generates business strategies, distributes thousands of direct purchasers, with more than 50 per cent operational efficiency; ADP leads AI agent produces real-time commodity referral and communication techniques, and A/B tests show a 26 per cent increase in sales orders.
ii. Mixed Hy4 will be added to visual understanding, Flash will play value for money, Agent Runtme will upgrade simultaneously
Then we say the second word: “efficient”.
As Agent continues to infiltrate the office and business landscape and begins to support enterprise-level production, there is an urgent need for more sophisticated tokens resource management to reduce costs, as the use and cost of Tokens in enterprises increases. Moving from using the strongest and most expensive models in rough use to pursuing value for money, matching different models with precision and performing different tasks.
Tencent clouds have an Agent infrastructure with high performance and very high value for value, stable isomeric calculators, working in depth with multiple chip manufacturers to construct new ecology of the national production capacity.
Under global stress, Tencent cloud depths fit the chips of multiple manufacturers, continuously optimizing the reasoning performance of the national chips, achieving good value for money and making them widely used in manufacturing industries such as smarts. Tencent clouds worked closely with mixed teams to boost five times the amount of single card through which a national chip was swallowed two weeks before it went online; we also helped cars to build the international head Tier1, a cluster of land landings of kilocalories, to support the scalable training of end-to-end intellectual driving LLM.
In the model, the Tencent hybrid Hy continues to evolve in the real business landscape. Over the past year, Hy has rebuilt the data and training infrastructure, an average of two months, with a large version of publish. Hy4 preview has been positioned as “living for productivity” and has been very stable and reliable in real productivity tasks such as code, office, science, and has been at the forefront of OpenRouter since publish, supporting stable services for beauties, WorldBuddy and others.
The hybrid LLMHy4 soon met with publish, a significant overall improvement compared to Hy4 preview, which strengthened the capacity of various long-range and real-productivity tasks, significantly increased the efficiency of the thinking chain and introduced visual understanding for the first time.
Hy4 flash will also provide, in the near future, publish, a practical model of extreme value for money that will evolve through the depth of Co-Design ' s high frequency in models and products, based on the ability to inherit such powerful capabilities as Hy4 vision and Agent.
Tencent has consistently pursued an open-source strategy based on the TokenHub platform, whereby users can call from the API to mainstream sources LLM such as hybrids, DeepSeek, GLM, Mini Max and Kimi. In conjunction with Token Plan ' s unified billing service, an enterprise can switch costs at very low cost and choose between multiple models on an as-needed basis.
In the content production scene, we also launched the Tencent cloud video AI WAND family, which provides the mainstream multi-modular model, the WAND-Vega multi-modular application model, and more than 100 self-study media processing tools, covering 90 per cent of the head-run short opera clients, producing 4.5 million pictures and 3 million seconds of video daily.
In addition, in the case of Agent Infra, we have today completely upgraded Agent Runtime, with systematic enhancements around key links such as context management, tool call, safety sandboxes, observations, assessments, etc., to make Agent more stable, manageable, more comprehensive and more expensive, and my colleagues will give a detailed presentation later.
Interception of high-risk operations, recording of call processes, enterprise data support encryption and recovery
The last key word is “more reassuring”.
As Agent deploys more and more within the network, both Agent and LLM will be exposed to more enterprise-level data and secure networks, operating environments and data will become particularly important. Enterprises must prevent not only the leakage of sensitive data to third parties, but also malicious attacks by internal and external users, as well as intellectual property protection and data theft beyond permission.
AI also magnifies the capabilities of the bad, and the threshold for Agent to commit evil has been significantly reduced. Recently, we saw a surge in the number of corporate key thefts, and later found the problem of repairing our security team and deploying a new generation of safety products to enhance protection.
Tencent itself operates many basic services at the national level, and we know that security is the foundation of trust, so it is always the primary consideration of products and services. For Agent, we are committed to building a “credible AI” governance system, consisting mainly of three aspects.
One is behavioral control. Agent represents who, what he can see, what he can do, and defines what he can do first; he can monitor, intervene, stop. In WorkBuddy, hidden malicious commands are identified and cleaned, extended capabilities are pre-mounted through security verification, high-risk operations such as data, over-the-counter access are deleted and undetermined operations are directly intercepted and confirmed by users.
Second, the process is traceable. Each Agent has an independent identity, incorporated into the Enterprise Unified Account and permission system; has seen data, used tools, executed actions, left marks throughout the journey, and administrators can locate problems by person, by time, by Agent, or uniform configuration rules to limit data outage.
The third is that data is manageable. The most unsure of enterprises is data: where is it, how is it spread, how is it deleted? Our answer is full-scale encryption, bottom-up recovery: private data are encrypted from terminal, transmission to storage, and sensitive missions support classified calculations; in case of error, overlay or destruction, document-level recovery is possible, so that an accident does not turn into irreparable loss.
These capabilities are also widely recognized. WorkBuddy was the first product in the country to be assessed through the CICA “Credible AI” and received the CCIA (China Cybersecurity Industry Union) “Personal Information Protection Top Evaluation”. At the same time, as a core unit, we have been involved in Shenzhen ' s AI agent Standard System of Artificial Intelligence, which provides credible AI best practices for standard-setting. These security capacity-building efforts have also won the endorsement of clients of political enterprises such as Guangdong Province, Shenzhen Airlines and Shenzhen.
We have always believed that innovation can only go further if security does not lie in efficiency. Tencent will continue to make good products and to build trust and become a trusted and trusted AI partner for the enterprise.
IV. A triple-digit increase in offshore AI demand to accelerate data centre layout and production by Agent Hey.
All right, let's finish the whole scene of Tencent AI, be more efficient and more reassuring. Next, I would like to share with you the progress and reflection of Tencent clouds in the service of globalization.
Tencent is also active in bringing home the home landing experience of LLM and AI Agent to the global level, ranging from arithmetical base, modelling services, to Agent-driven technology platforms and industry solutions, helping more global firms build high-value, secure and manageable AI technology depots.
Over the past year, the number of active overseas clients of Tencent Cloud International has doubled, covering over 150 countries and territories worldwide. In several key markets, we're all out of business:
In Indonesia, we signed and actively promoted the three local operators, XLSMART, Telkomsel and Indosat, in full cooperation with AI in the cloud.
In Thailand, we continue to reduce efficiency gains for more business blocks for large and active groups, accelerate the intellectual transformation and enter head operators, exchanges and banks.
In Hong Kong and Macao, we have served clients such as Hong Fu, the Government Digital Policy Office, and business has grown steadily for the seventh consecutive year.
In Europe, we have continued to expand our in-depth cooperation with global Internet platform enterprises such as iyzico, eMAG.
Looking to the future, AI is reshaping the global cloud market logic of choice, with clients choosing the cloud, moving away from resources, costs and stability, to modelling capacity and the AI tool chain.
China ' s AI has a leading position in terms of industry landscapes, product experience and value for money. Tencent clouds are bringing these capabilities to the global level, helping to reach the sea and serving global clients.
First, we will accelerate the construction of a global data centre.
The rate of Tencent cloud openings abroad has long been industry-led, with over 2 million servers deployed globally. Over the past year, we have added or expanded data centres in Frankfurt, Malaysia, Mexico, Osaka, Riyadh and Thailand. In the coming year, we will also further expand the geography of Asia, the Middle East, Europe and the Americas.
Second, it is to build high-value infrastructure.
In the face of the surge in computing and storage needs brought by the great wave of AI, we continue to design the best software and hardware architecture for our clients. For example, Starfish's latest flagship fleet, supporting 1024 nuclears, cloud-backed cloud servers upgraded to the latest SA10, together with extreme memory management techniques, provide customers with extremely cost-effective computing platforms.
In the face of rising storage costs, we have also designed stratification storage programmes for many clients; the current COS tape storage cumulatively exceeds 2,000 pB, helping global customers to file “extreme cold data” at low cost.
For clients with their own data centres and upgrading from virtualization technology to the cloud core, specialized cloud TCE deployments have been reduced from 14 to 8, further reducing deployment thresholds and costs.
In addition, we are keen to seize the opportunities of AI applications to enhance global delivery capacity.
Over the past year, the overseas AI demand for Tencent clouds has grown in three digits, and global businesses have a high level of interest in and demand for our country ' s AI products ecology.
At the model level, the open-source model, with high-frequency scenarios such as office, knowledge question and answer, and content generation, presents a good experience with a very high aggregate price ratio, with Hy4 Preview coming online and soon receiving advice from many overseas enterprises. In the future, we will also invest more resources in securing high-quality Token global supplies.
At the product level, we are accelerating access to all categories of Agent, such as CodeBuddy, WorkBudy, Tencent, Miora, ADP, MAGIC AI, Agent Runtme, and so on, so that both the expatriate staff of Chinese enterprises and local teams of overseas clients can work together on the same set of AI workshops.
In terms of servicing overseas clients, we have continued to assist overseas platform enterprises in building local Superapp ecology, upgrading the small program platform programme Superapp as a Service with CodeBuddy, and producing small programs by developers in natural languages. Based on it, the Indonesian indigenous operator Telkomsel has doubled in a year to construct a small program ecology.
On audio and video scenes, during this year ' s USA World Cup, WAND audio and video AI technology provided technical services to official transmission platforms in 17 countries and territories. During the Nagoya Games, the IOC President gave live praise to China's rebroadcast technology, which we created together with OCSS.
In the cloud migration scene, we have developed more than 20 relocations of Skills based on CodeBuddy and WorldBuddy, which took only 4.5 months to help the Indonesian head operator, XLSMART, to make the migration smooth, while also developing more than 200 localization functions.
Today, the global competitiveness of Chinese science and technology enterprises comes not only from costs but also from the combined capacity of AI, clouds and engineering. Tencent clouds hope to be a bridge between China's AI and global industries, helping more firms grow in the global market.
Guests, partners, look back to the development of information technology and real industrial change, not only from technological breakthroughs, but also from deep technological operations, strengthening upstream and downstream connectivity and even changing organizational patterns.
So does AI. From “use it” to “use it well”, the key is to get AI into the process, create productivity and continuously improve business efficiency and investment returns. Tencent will insist on long-term input, continuous grinding of products, open platform capacity, and support to enterprises in building an operational and evolutionary Agent productivity system.
In the process, we have always opened our doors to our partners and looked forward to working with you to build a prosperous Agent industrial ecology.
Thank you.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.