10 月 22 日至 24 日，由极客邦旗下 InfoQ 中国主办的 QCon 全球软件开发大会 2026 在上海举行。来自开发、架构、技术管理和产业一线的 1200 余位从业者齐聚现场，围绕 AI 技术演进与企业落地展开交流。
本次大会邀请到来自阿里巴巴、字节跳动、腾讯、快手、蚂蚁集团、小红书、美团、飞猪、上海交通大学、月之暗面、无问芯穹、网易、记忆张量、货拉拉、香港大学、香港科技大学、阶跃星辰、亚马逊云科技、华为、小影科技、平安科技、蔚来、涛思数据、观测云、焱融科技、43 AI、ATEVE、Datus.AI、Dexmal、Frontier、微软、蚂蚁灵波科技、Mooncake 社区、Muka Robotics、NoDesk AI、XTransfer、zoowork.ai、上汽集团、中国计算机行业协会网络和数据安全专业委员会、中泰证券、云深处科技、京东、优码云、传音控股、COSMOPlat、去哪儿、哔哩哔哩、商汤、大晓机器人、小盾未来、平安人寿、携程、智元、江苏移动、清华大学、爱奇艺、生数科技、百度等共 70 家企业与机构的 140 位专家与实践者同台分享。
在 2026 年 QCon 全球软件开发大会（上海站）主会场，从技术趋势到企业落地，从办公方式到工业数据，再到算力基础设施和内容应用，六位技术实践者沿着一条清晰路径，回答了同一个问题：AI 如何真正进入生产系统。
从 Agent 到自主智能
上午第一场演讲中，极客邦科技总编赵钰莹以《从 Agent 到自主智能：2026 AI 技术演进与 2027 趋势前瞻》为题，回顾了 AI 从模型能力竞争走向 Agent 工程化的过程。
在她看来，随着基础模型能力不断提升，行业关注点已经从模型本身扩展到 Agent 工程、多智能体协作、上下文管理、评测体系和生产可靠性。
Agent 不再只是能够回答问题的助手，而是开始承担任务拆解、工具调用、过程执行和结果交付。但能力越强，系统复杂度和潜在风险也会同步增加。
赵钰莹指出：
“模型能力并不是决定 Agent 落地效果的唯一因素，工程框架同样会直接影响任务完成质量。”
她特别提到，多智能体并不意味着 Agent 越多越好。过多的 Agent 可能带来更高的通信成本、错误传播风险和协作损耗。企业真正需要关注的，是任务是否完成、结果能否验证、过程能否追踪，以及失败后能否恢复。
AI 带来的变化也不只发生在技术系统内部。随着 Agent 逐渐进入业务流程，企业原有的岗位边界、协作方式和决策机制都可能发生变化。
“AI 转型真正深刻的变化，不只是效率提升，而是岗位分工、业务流程和组织结构的重新设计。”
这意味着，企业下一阶段的 AI 改造不能只是研发团队的工作。业务专家也需要参与 Agent 目标、业务上下文、执行规则和验收标准的定义。只有技术团队与业务团队共同参与，AI 才可能从一个演示能力，转变为稳定的生产能力。
助力企业 AI 加速落地：从知识服务到供需大脑
第二场演讲中，极客邦科技创始人兼 CEO 霍太稳围绕《助力企业 AI 加速落地》，分享了极客邦科技在企业 AI 服务方向上的新思考。
今年是极客时间成立 9 周年。9 年来，极客时间持续通过课程、内容和社区连接技术人才与企业。面对 AI 带来的变化，极客邦科技将进一步把知识、人才与生态资源连接到企业 AI 落地实践中，并将公司使命升级为“助力企业 AI 加速落地”。
“AI 落地卡住的地方，从来不是技术，是信任和信息。”
使命升级：从知识服务与人才培养，走向企业 AI 实践
在本届 QCon 上海期间，极客时间围绕 AI 落地推出了多项内容与服务，包括目标汇集 100 个真实业务实践的《AI 落地案例集》，以及覆盖大模型通识、Agent 工具、工程化和面试方向的 18 个 AI 工程师知识库。
模力工场升级：打造企业 AI 落地的“供需大脑”
依托服务 4000 多家企业、连接 400 多万用户的积累，极客邦科技发现，企业和 AI 交付团队正面临着相互对应的难题：企业难以从众多产品和供应商中完成选型，交付团队则常常面临获客难、建立信任难和获取高质量需求难。
“一边找不到靠谱的团队和产品，一边找不到靠谱的客户和需求。”
因此，模力工场从 AI 社区平台升级为企业 AI 落地服务平台，围绕企业需求发布、场景诊断、方案匹配、FDE 交付和项目协同展开，并逐步沉淀场景库、人才库、解决方案库和项目案例库，建设评价与信用体系。
霍太稳表示，极客邦科技近 20 年来持续连接企业与技术人才，InfoQ 关注技术演进与行业实践，TGO 鲲鹏会连接科技领导者，极客时间参与人才培养。这些积累让极客邦科技既能理解企业需求，也能连接具备实践能力的人与团队。
“这不是从零开始，是水到渠成。”
面向企业，模力工场希望提供从场景诊断、方案共创、FDE 落地实施到价值优化的完整路径；面向生态伙伴，则通过联合方案、客户共创与交付协作，让优质产品、技术和方案进入真实业务场景。
从知识获取、人才培养，到方案匹配和项目交付，极客邦科技正在形成一条从技术内容到企业 AI 落地的完整链路，让好产品找到场景，让真实需求找到答案。
软硬一体：AI 正在重新定义企业的工作方式
在第三场演讲中，阿里巴巴千问办公副总裁任卿围绕《软硬一体：打造 AI 时代的工作方式》，讨论了企业如何借助 AI 重构日常工作流程。
任卿认为，企业级 AI 与通用对话产品的差别，首先在于它面对的不是一个孤立问题，而是一个包含组织关系、业务流程、权限体系和历史信息的复杂工作环境。
模型可以沿着公共数据和公共知识不断进化，但企业真正独特的资产，往往存在于自身的组织关系、业务流程、客户信息、产品规则和历史经验之中。
这些信息能否被准确组织、调用和治理，决定了 Agent 是否真正理解企业。
任卿将这一能力概括为上下文工程：
“对企业级 AI 来说，上下文工程带来的差别不是一百倍，而是零和一，是可用与不可用的区别。”
上下文并不只是一段提示词，还包括用户身份、权限范围、历史行为、业务对象及其关系。只有把这些信息组织成 Agent 能够理解和调用的结构，AI 才可能从通用助手进一步成为企业工作伙伴。
但 Agent 越接近真实业务，风险也越高。能够读取数据、调用系统和执行操作的 Agent，一旦判断错误，影响可能远大于一次错误回答。
因此，企业级 Agent 不仅要有“加速能力”，还要有“刹车能力”：在关键操作前确认权限，对敏感动作进行留痕，并在异常情况下停止、回滚或转交人工处理。
对企业而言，AI 时代的工作方式并不是简单地把原有流程自动化，而是重新思考哪些工作应该由人完成，哪些工作可以交给 Agent，以及人和 Agent 如何共同承担责任。
打造工业本体：让工业数据成为 AI 可以理解的基础设施
企业 AI 进入工业场景后，首先要面对的往往不是模型能力，而是数据语义问题。
涛思数据创始人兼 CEO 陶建辉在《打造工业本体，构建 AI 原生的工业数据基础设施》中指出，工业企业并不缺少数据。SCADA、设备系统、时序数据库和业务系统中积累了大量信息，但这些数据长期分散在不同系统中，缺少统一的语义表达。
一个温度、压力或转速数据，如果缺少设备对象、采集位置、计量单位、时间范围和工况信息，对 AI 来说只是一个孤立数字。
陶建辉表示：
“只有把一个数值对应的上下文完整提供出来，这个数据才有意义；否则 AI 只能靠猜。”
因此，工业 AI 首先要解决的不是“让模型更聪明”，而是建立能够被人和 AI 共同理解的数据结构。
涛思数据正在工业数据基础设施之上构建工业本体层，用于描述设备、属性、事件及其关系，为跨系统查询、故障分析和 Agent 调用提供稳定的语义基础。
陶建辉同时强调，AI 并不需要替代所有确定性软件。
对于规则明确、结果稳定、需要高频执行的任务，更适合让 AI 生成经过验证的程序，再由系统持续运行，而不是每次都让模型重新推理。
“可以用 AI 生成确定性的程序，但没有必要让 AI 每次都重新执行确定性的逻辑。”
在这一分工下，AI 更适合处理开放问题、复杂意图和动态决策，传统软件则负责高可靠执行。两者结合，才能兼顾灵活性和稳定性。
对于缺少资金、人才、专家和时间的中小企业而言，工业 AI 的价值不只是提供一个新工具，更是帮助它们将分散的数据转化为可理解、可调用、可持续利用的业务资产。
Agentic Infra：让基础设施支撑 Agent，也让 Agent 改造基础设施
当前智能体正推动全球 Token 需求集中释放，并呈现长期持续的指数级增长趋势。上海交通大学人工智能学院副教授、无问芯穹联合创始人兼首席科学家戴国浩在《以 Agentic Infra 技术实践支撑词元工厂持续演进》中表示，芯片、网络和数据中心等物理设施的供给受到制造能力、电力及配套条件约束，仅靠计算资源规模扩张，很难满足指数级增长的 AI 需求。要弥补这一供需缺口，一方面需要以更少的能源和算力投入生产更多 Token，另一方面也要提升 Token 向实际任务成果的转化效率。两者共同决定了 AI 基础设施的有效产出能力。
戴国浩提出，无问芯穹的 Agentic Infra 包含两个方向：
“不仅是 Infra for Agent，还有 Agent for Infra。”
一方面，基础设施要稳定支撑 Agent 的运行与进化；另一方面，Agent 也可以参与算子优化、系统运维、资源调度和基础设施自身的演化。AI 基础设施面临的机会不再只有“如何更高效地生产 Token”，还包括“如何利用 AI 提升整个生产系统的效率”。
过去，企业更关注单卡吞吐量、Token 价格和响应延迟；未来则需要进一步关注：完成一个可验证任务，需要多少综合成本。
真实任务通常不是一次模型调用，而是一条包含规划、执行、工具调用、结果检查和失败重试的工作流。如果任务频繁失败或需要反复调用，即使单次推理价格很低，整体成本仍然可能很高。
因此，Agentic Infra 的目标不只是让系统运行得更快，也包括让系统能够持续优化任务路径，并不断提高每个 Token 所产生的实际价值。
戴国浩提到：“Agent 应用中，模型能力固然重要，但模型并不是决定最终任务效果的唯一因素。Agent 效果还取决于 Harness 的组织方式。”一系列已有工作已证明，智能系统的改进对象，正从生成轨迹扩展到 Workflow 与 Agent 自身。
围绕这一技术方向，戴国浩分享了在自动算子优化、科学推理和具身智能端侧推理部署等场景中的实践案例。在这些场景中，Agent 不再只是上层应用的调用者，而是深入到了底层技术系统的优化环节。他表示：“无问芯穹相信能够建立一套 Token maxxing 的 Agent Harness，在满足质量约束的前提下最小化端到端成本，通过优化轨迹数据进一步改进 Agent 系统的优化策略，逐步向 RSI 推进。”
从“生产更多 Token”拓展到“以更低成本交付更多有效任务成果”，AI 基础设施正向着能够持续交付有效 AI 生产力的新型供给体系逐步演进。
从工具到 Agent：内容创作产品的三次范式迁移
最后一场演讲中，小影科技创始人兼 CEO 韩晟以《从工具到 Agent：小影内容创作的三次范式迁移》为题，分享了 AI 如何改变内容创作产品和组织方式。
在移动互联网阶段，产品主要向用户提供剪辑、模板等创作工具；进入生成式 AI 阶段，图片、视频、音乐等素材可以由模型生成，但用户仍然需要自己组织完整的工作流；到了 Agent 阶段，产品开始从提供工具转向理解用户意图并交付最终结果。
同样一句“制作一条宣传视频”，在不同用户和场景中，背后的目标可能完全不同。美甲店希望吸引顾客到店，奶茶店可能关注促销转化，电商商家希望突出产品卖点，自媒体创作者则可能更关注内容完播率。
因此，Content Agent 需要理解的不是一句指令，而是用户、场景、目标和结果标准。
产品形态变化的同时，企业内部的技术架构和协作流程也需要同步调整。
韩晟指出：
“如果流程不改变，只是让每个岗位在自己的流程里面提效，整体开发效率还是不高的。”
小影科技的实践是，通过 AgentOps 统一管理模型路由、Skill、工作流、评测、安全和可观测能力，让业务团队可以组合已有能力，快速搭建产品原型，并在真实用户环境中验证结果。
在这一模式下，技术团队不再只是逐项实现业务功能，而是更多负责建设平台、组件和运行机制；业务团队则基于这些组件快速构建 Agent、验证需求，并根据用户行为持续修正成功标准。
真正需要长期沉淀的，也不只是对某个模型的调用，而是企业对用户的理解、对结果的定义，以及围绕真实反馈形成的数据闭环。
从能力展示到任务交付，企业 AI 进入生产化阶段
按照当天主会场的演讲顺序，六位嘉宾分别从趋势、企业服务、工作方式、工业数据、基础设施和内容应用展开讨论，最终形成了一条完整的技术与产业路径：
AI 首先需要理解企业的业务上下文，随后依赖数据语义和基础设施完成任务执行，最终通过工作流、评测和组织协作，将模型能力转化为可验证的业务结果。
这也意味着，企业 AI 的竞争正在发生三重变化：
第一，评价标准从模型能力和功能数量，转向任务完成率、结果可验证性和业务价值。
第二，企业需要沉淀的核心资产，从一次性的模型调用，转向数据语义、上下文结构、评测标准和用户行为反馈。
第三，AI 带来的组织变化不再局限于研发部门。产品、业务、技术、测试和运营团队，都需要围绕任务结果重新组织协作方式。
对企业而言，接入大模型只是开始。真正的挑战，是建设一套让 AI 能够被使用、被验证、被复用，并持续创造业务结果的生产系统。
分会场全景：沿着 AI 生产化路径展开
如果说主会场讨论的是 AI 正在走向任务交付，那么分会场则进一步回答了：一套真正可用的 AI 系统，需要经过哪些工程环节，才能进入生产环境。
本届 QCon 上海的分会场内容，可以概括为一条从系统构建到产业落地的完整路径。
先解决系统如何构建
AI Native 架构、企业级多 Agent 架构与系统、Agent as a Service、Loop Engineering 等专题，聚焦 Agent 的架构设计、任务编排、状态管理、工具调用、执行环境和安全边界。
与此同时，Agent 自主进化：从记忆到持续学习进一步讨论了 Agent 如何管理长期记忆、积累执行经验，并在反馈中持续改进。
再解决系统如何稳定运行
当 Agent 进入生产环境，可靠性和可控性成为新的重点。AI Infra：算力效率决定规模化落地关注训练、推理、调度和软硬件协同；AI Agent 可观测与持续改进工程讨论如何通过 Trace、评测和反馈，让 Agent 的行为可见、结果可测、系统可持续优化。
理性驾驭 AI、Harness、Loop 与 Graph：Agent 产品设计的演进之路、AI 时代的攻防重构以及驾驭 Agent：记忆、观测、沙箱与具身实践等专题，则进一步覆盖 SRE、运行安全、沙箱隔离、风险治理和 Agent 产品设计。
AI 正在重构软件研发和数据系统
在研发环节，AI 驱动的软件工程和 Vibe Coding 时代的新质量债关注 AI Coding、研发 Agent、代码评审、测试验证和可信交付，讨论如何解决“代码生成更快，但质量验证更难”的新问题。
在数据领域，Agentic 时代的数据系统重构聚焦 Data Agent、数据语义层和 Context Layer；端云协同与 AI Native 终端生态：从端侧模型到智能体应用则讨论端侧模型、异构计算、端云联合推理以及跨设备智能体体验。
从技术系统走向真实产业
AI 的价值最终需要在真实业务中验证。数据驱动的具身智能：迈向规模化之路关注机器人数据闭环、物理智能和规模化落地；千行百业 Agent 创新实践覆盖客服、营销、交易和企业生产等应用场景。
AIGC 内容工业化聚焦内容生产、质量评估、安全合规和多 Agent 协同；金融 AI Native 工程实践：从研发提效到业务破局则将视角延伸至保险、证券、跨境支付和金融风控等场景。
技术变化也正在改变组织和人的角色
除了技术专题，大会还设置了打造 AI Native 型组织和技术人的下一站，讨论企业如何重构协作机制、工程管理和人才能力，以及技术人如何从代码实现者走向系统设计者和任务结果负责人。
此外，大会还设置了 KIRO「鬼点子计划」拍卖会、AI 落地见真章——极客时间 9 周年×QCon 上海同行之夜和把 AI 这道题从头做一遍等特色活动，为参会者提供动手实践、真实案例讨论和面对面交流的机会。
从架构、基础设施和数据系统，到研发交付、组织变革和行业应用，本届 QCon 上海的分会场共同呈现了一张 AI 生产化路线图：AI 正在从一个需要被调用的能力，逐渐成为可以被设计、运行、评测和持续改进的生产系统。
生态共绘：展位交流与赞助支持
在大会现场，展区同样有人气十足的技术交流。亚马逊云科技、TDengine（涛思数据）、观测云、合合信息、焱融科技、智元 AIMA 开发者社区、AReaL、IPIP 等企业围绕 AI 基础设施、智能体开发平台、数据与算力及工程化工具，展示前沿技术与实践方案，并与参会者展开深入交流。腾讯云等合作伙伴也通过赞助支持和生态协同，共同参与本届大会。
此外，中欧国际工商学院 EMBA、独角兽创业营、TGO 鲲鹏会作为学术合作单位，为大会提供支持。多方伙伴共同连接技术创新与产业实践，推动 AI 技术在真实业务场景中持续落地。
热度延续：12 月 AICon 北京站见
12 月 18–19 日，AICon 全球人工智能开发与应用大会将在北京举办。本届大会以探索 A 应用边界为主题，设置 13 大技术专题，覆盖大模型后训练、推理优化、Agent 工程、知识工程、智能体评测与安全、自进化系统、AI 软件交付、原生架构及组织变革等核心方向，汇聚一线技术专家，深入分享 AI 技术突破与企业规模化落地的真实实践。
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。
From 22 to 24 October, the QCon Global Software Development Conference, hosted by InfoQ China under the banner of Polar State, was held in Shanghai. More than 1200 practitioners from the development, architecture, technology management and industry sectors gathered on the ground to communicate with businesses around the AI technology evolution.
The Congress invited REDnote, United States Corps, Flying Pigs, Shanghai Transport University, Dark Sides of Moon, No-Quest Dome, Memories, Melala, University of Hong Kong, Hong Kong University of Science and Technology, Step Star, Amazonian Cloud Science, Huai, Small Shadow Technology, Safe Science, Indemnity, Toss Data, Observation Clouds, Melting Technology, 43 AI, ATEVE, Datus.AI, Dexmal, Frontier, Microsoft, Questing Technology, Mooncake Community, Muka Robotics, NoDesk AI, XTransfer, Zoworkai, Shanghai Group, China Computer Industry Association Network and Data Security Committee, China Securities, Yuntai, Kyotung, Euntung, Voice Control Unit, COSMOPlat, where to go, Zoncake Community, Chongtung, Da Xiaohuatung, Xiaocai, Tsun, Zhi Tsun, ZhiQui, ZhiQu, ZhiQu, ZhiQu, ZhiQu,
In 2026, in the main venue of the QCon Global Software Development Conference (Shanghai Station), from technological trends to business landings, from office patterns to industrial data, to computing infrastructure and content applications, six technical practitioners followed a clear path and answered the same question: how to get real access to production systems.
From Agent to autonomous intelligence
In the morning ' s first session, Zhao Zhao, editor-in-chief of technology in the state of Polarism, reviewed the process of AI moving from a modeling competition to Agent engineering, under the theme " From Agent to Auto-Intelligence: Technology Evolution and Trends 2026 AI ".
In her view, as basic modelling capacity had increased, industry concerns had been expanded from the model itself to the Agent Project, multi-AI agent collaboration, context management, evaluation systems and production reliability.
Agent is no longer merely an assistant who can answer questions, but begins to take on tasks such as dismantling, tool call, process execution and delivery of results. However, the greater the capacity, the greater the complexity of the system and the increased potential risks.
Zhao Zhou Zhou said:
“Model capabilities are not the only factor determining the impact of Agent's landing, and the engineering framework also has a direct impact on the quality of mission completion.”
She mentioned in particular that multi-AI agent did not mean that the more Agent the better. Too many Agents can lead to higher communication costs, mistransmission risks and collaborative losses. The real concern of enterprises is whether the task is completed, whether the results are validated, whether the process is tracked and whether it can be restored after failure.
AI brings about changes that are not only within the technology system. As Agent moves into business processes, changes in the existing job boundaries of the enterprise, the modalities of collaboration and the decision-making mechanisms are possible.
“AI A truly profound change in transition is not just an increase in efficiency, but a redesign of the division of labour, business processes and organizational structure.”
This means that the next phase of the AI transformation of an enterprise cannot be the work of a research and development team. Operational experts also need to be involved in the definition of Agent objectives, operations context, enforcement rules and acceptance criteria. Only when technical and business teams are involved can AI move from a demonstration capacity to a stable production capacity.
AI Accelerated landing: from knowledge service to supply and demand brain
In the second presentation, the founder and CEO of Polar Science and Technology was able to share new ideas on the direction of Polar Science and Technology in the direction of business services.
This year marks the ninth anniversary of the Holloway. Over the past nine years, it has been through courses, content and community connections between skilled people and businesses. Facing the changes brought about by AI, Polar Science and Technology will further link knowledge, talent and ecological resources to enterprise AI landing practice and upgrade the corporate mission to “Accelerator Enterprise AI Accelerator landing”.
“AI places where land is stuck, never technology, trust and information.”
Mission Upgrading: from Knowledge Services and Talent Development to Enterprise AI Practice
During the current QCon Shanghai period, a number of content and services were launched around AI landing, including the " AI landing case collection ", which brings together 100 real business practices, and the 18 AI engineers knowledge base covering LLM literacy, Agent tools, engineering and interview directions.
Modular workshop upgrading: building the firm AI "supply and demand brain"
With more than 4000 businesses, more than 4 million users connected, Grameen Technologies found that companies and AI delivery teams are facing each other: It is difficult for enterprises to choose from a wide range of products and suppliers, and delivery teams often face difficulties in gaining access, building trust and obtaining high-quality demand.
“There is no reliable team or product on the one hand and no reliable client or demand on the other.”
As a result, the Modular workshop was upgraded from an AI community platform to an IAI landing service platform, which built an evaluation and credit system around business needs publish, scene diagnostics, matching programmes, FDE delivery and project synergies, and gradually settled the landscape, talent, solutions and project case bank.
According to Houtha, the technology of Monopoly has been connecting firms and technologists for almost 20 years, the InfoQ focuses on technology evolution and industry practice, and the TGO luminum connects technology leaders to participate in talent development. These accumulations allow the technology to understand business needs as well as to connect people and teams with practical skills.
"This is not from scratch, it's from water to canals."
Business-oriented, model workshops seek to provide a complete path from scene diagnosis, co-generation, FDE implementation to value optimization, and eco-partner collaboration through joint programmes, client co-generation and delivery to bring quality products, technologies and programmes into the real business landscape.
From knowledge acquisition, talent development, to programme matching and project delivery, polar state technology is forming a complete link from technology content to enterprise AI landing, where good products can be found and real needs answered.
Soft and hard: AI is redefining the way an enterprise works
In the third presentation, Alibaba asked the vice president about how companies could use AI to re-engineer the daily workflow process around " Soft and hard: how to build the AI era ".
According to the Secretary-General, the difference between a company-level AI and a generic dialogue product lies first and foremost in the fact that it is not facing an isolated problem, but rather a complex work environment encompassing organizational relationships, business processes, permission systems and historical information.
Models can evolve along public data and knowledge, but the real unique assets of enterprises are often found in their organizational relationships, business processes, customer information, product rules and historical experience.
Whether this information is accurately organized, mobilized and governed determines whether Agent really understands the business.
He summarized this capability as context:
“For Enterprise AI, the difference made by the context project is not 100 times greater, but zero and one, and is the difference between available and unavailable.”
context is not just a section of prompt, but also includes the user identity, the scope of permission, historical behaviour, the object of the business and its relationship. It is only by organizing this information into a structure that Agent can understand and call on, that AI can further become a business partner from the General Assistant.
But the closer Agent is to real business, the higher the risk. Agent, who can read data, call the system and execute the operation, may have a much greater impact than a wrong answer if the judgement is wrong.
Thus, at the enterprise level, Agent has not only an “acceleration capability” but also a “brakes capability”: confirming permission before critical operations, leaving marks on sensitive movements, and stopping, rolling back or transferring manual processing in exceptional circumstances.
For businesses, the AI-era approach is not simply to automate old processes, but to rethink what should be done by people, what can be done by Agent, and how people and Agent can share responsibility.
Building an industrial home: making industrial data an understandable infrastructure
Enterprise AI, when entering the industrial scene, is often confronted first and foremost not with modelling capacity, but with data semantics.
The founder and CEO of Toth data, in Building the Industrial Base, Building the Initial Industrial Data Infrastructure, noted that industrial enterprises did not lack data. The SCADA, equipment systems, time series databases and operating systems accumulate a large amount of information, but these data are scattered over a long period of time across different systems and lack a uniform syntax expression.
A temperature, pressure or rate of rotation data is an isolated number for AI if the equipment object, location, unit of measure, time range and condition information are missing.
Ta Jianhui says:
"This data will only be meaningful if the full context of a value is provided; otherwise AI can only guess."
So, industry AI first has to deal not with “making models smarter”, but with creating data structures that can be understood in common with AI.
Tothic data are building an industrial infrastructure above the industrial data infrastructure to describe equipment, properties, events and their relationships, providing a stable semantic basis for cross-system queries, failure analysis and Agent calls.
Tao Jianhui also stressed that AI did not need to replace all certainty software.
For missions with clear rules, stable results and high frequency (HF) performance, it is more appropriate for AI to generate a validated program and then run on a continuous basis by the system instead of having the model re-engineer each time.
“A procedure can be used to generate certainty, but there is no need for AI to re-exercise the logic of certainty each time.”
Under this division of labour, AI is better placed to deal with openness issues, complex intentions and dynamic decision-making, while traditional software is responsible for high-level and reliable implementation. In combination, flexibility and stability can be balanced.
For SMEs that lack finance, talent, expertise and time, the value of AI is not only to provide a new tool, but also to help them translate dispersed data into understandable, accessible and sustainable business assets.
Argentina Infra: Making infrastructure support Agent, and making it work.
The current AI agent is contributing to a concentrated global token demand release and is showing a long-term trend of exponential growth. The Associate Professor at the College of Artificial Intelligence of Shanghai University of Transport, Co-founder and Chief Scientist of the Unquestioned Cricket Dome, Dai Kuo-Ho, in his “Continuing the evolution of the meta-factor with the Agenic Infra Technical Practice”, stated that the supply of physical facilities such as chips, networks and data centres is constrained by manufacturing capacity, electricity and the conditions accompanying them, and it is difficult to meet the AI demand for exponential growth simply by calculating the scale of resource expansion. To fill this gap between supply and demand, on the one hand, and production of more Token with less energy and capital, on the other, and token, on the other hand, to increase the efficiency of its transformation to the results of the actual mission. Together, they determine the effective output capacity of AI infrastructure.
Deok-ho suggested that the no-questioned Agenic Infra contains two directions:
"Infra for Agent, and Agent for Infra."
On the one hand, infrastructure needs to be stable to support the operation and evolution of Agent; on the other hand, Agent can also participate in the optimization of algorithms, systems, resource mobilization and the evolution of infrastructure itself. AI opportunities for infrastructure are no longer limited to “how to produce more efficiently”, token, but also to “how to use AI to improve the efficiency of the entire production system”.
In the past, enterprises have focused more on single-calorage, token prices and response delays; in the future, further attention will be needed on how much of the combined cost is required to complete a verifiable task.
real task is not usually a model call, but a workflow containing planning, execution, tool call, results check and failure retry. If missions fail frequently or need to be called repeatedly, the overall cost may still be high, even if the price of a single reasoning is low.
Therefore, the goal of Agenic Infra is not only to make the system work faster, but also to allow the system to continuously optimize the mission path and continuously increase the real value generated by each Token.
Deok-ho mentioned: “Agent application, while model capabilities are important, models are not the only factor determining the effects of the final mission. The effect of Agent also depends on the way Harness is organized.” A series of established work has proven that the target for the improvement of the smart system is expanding from generating a trajectory to the WorldFlow and Agent itself.
In this technical direction, Dai Kwok-ho shared practical examples of scenarios such as autocalculator optimization, scientific reasoning and deployment of smart side reasoning. In these scenes, Agent is no longer just a caller for top-level applications, but goes into the optimization of the bottom technology system. He said: “The no-question core is confident of building a set of Token maxxing Agent Harness that minimizes end-to-end costs, subject to quality constraints, and further refines the Agent system's optimization strategy by optimizing trajectory data and moving towards RSI.”
From “more production token” to “more effective mission results at lower cost”, AI infrastructure is evolving towards new supply systems capable of delivering effective AI productivity on a sustainable basis.
Three paradigm migrations from tool to Agent: content creation
In the final presentation, CEO Han Xin, founder and founder of small film technology, shared AI on how to change the way content is produced and organized, under the theme " From tools to Agent: three paradigms for small film creation ".
During the mobile Internet phase, the product primarily provides users with creative tools such as clippings, templates, etc.; enters the generation AI phase, where asset can be generated by models, but users still need workflow, which is fully organized; to the Agent phase, the product begins to shift from providing tools to understanding user intentions and delivering the final results.
In the same sentence, “to produce a promotional video”, the objectives behind it may be completely different from those of different users and scenes. The nail shop wants to attract customers, the milk tea shop may be concerned with the promotion of the conversion, the electric dealer may want to highlight the selling point of the product, and the self-media creators may be more concerned with the content completion rate.
It is therefore not a directive that Content Agent needs to understand, but rather a user, scene, objective and outcome standard.
Changes in product patterns need to be accompanied by simultaneous adjustments in the technological architecture and collaborative processes within the enterprise.
Han Xin says:
“The overall development will not be efficient if the process is not changed and only if each post is made effective in its own process.”
The practice of small-scale video technology is to allow business teams to combine their capabilities to quickly build product prototypes and validate results in a real user environment through the AgentOps integrated management model routing, Skill, workflow, assessment, safety and observable capabilities.
Under this model, the technical team is no longer simply implementing business functions on a line-by-line basis, but is more responsible for building platforms, components and operating mechanisms; the operational team is based on these components for fast-track construction of Agent, validation needs and continuous revision of success criteria based on user behaviour.
What is really needed is long-term deposition, not just a call to a model, but an enterprise's understanding of users, definition of results, and data closure around real feedback.
From capability demonstration to task delivery, enterprise AI goes into production
In accordance with the speaking order of the main venue of the day, the six guests discussed trends, business services, working methods, industrial data, infrastructure and content applications, resulting in a complete technology and industry path:
AI first needs to understand the business of the enterprise context and then rely on data semantics and infrastructure to carry out its tasks, ultimately translating model capabilities into verifiable business results through workflow, evaluation and organizational collaboration.
This also means that the competition of the firm AI is changing threefold:
First, evaluation criteria move from model capacity and number of functions to mission completion rates, validation of results and operational values.
Second, enterprises need to settle core assets, moving from one-time models to data semantics, context structures, assessment standards and user behaviour feedback.
Third, the organizational changes brought about by AI are no longer limited to the R & D sector. Products, operations, technology, testing and operational teams need to reorganize collaborative approaches around mission results.
For businesses, access to LLM is only the beginning. The real challenge is to build a production system that allows AI to be used, validated, reused and continuously generate business results.
Division Panorama: along the AI production path
If the main session was about AI as it moved towards task delivery, the session was further answered: a truly functional AI system, what works were needed to enter the production environment.
This session of the QCon Shanghai branch can be summed up as a complete path from system construction to industrial landing.
First solve how the system is built.
AI Native Architecture, Multi-Enterprise Agent Architecture and Systems, Agent as a Service, Loop Engineering, etc., focused on the design, tasking, state management, tool call, implementation of environmental and security boundaries.
At the same time, Agent's autonomous evolution: from memory to continuous learning further discussed how Agent managed long-term memory, accumulated implementation experience and continued improvement in feedback.
Resolving how the system works steadily
When Agent enters the production environment, reliability and control become new priorities. AI Infra: Accomplishable efficiency decisions scale up attention to training, reasoning, scheduling and software and hardware synergies; AI Agent can observe and continuously improve the project to discuss how Agent ' s behaviour can be visible, measurable, and system sustainable and optimized through Trace, evaluation and feedback.
The topics of rational driving, the evolution of the design of AI, Harness, Loop and Graph: Agent products, the anti-restructuring of the AI era, and the management of Agent: memory, observation, sandboxes and hands-on practices further cover SRE, operating safety, sandbox isolation, risk management and Agent product design.
AI Reconfiguring software development and data systems
At the R & D stage, AI-driven software engineering and the new quality debt focus of the Vibe Coding era AI Coding, R & D Agent, code review, test validation and credible delivery discuss how to address the new problem of “more rapid code generation, but more difficult quality validation”.
In the area of data, the data system of the Age of Age of Age re-focuss Data Age, Data Semantic Layer and Context Layer; end clouds synergizing with AI Native terminal ecology: The application from end-side model to AI agent discusses end-side model, isomer computation, cloud-end joint reasoning and cross-equipment AI agent experience.
From technology systems to real estate
The value of AI ultimately needs to be validated in real business. Data-driven smartness: the path to scale focuses on robotic data closed loops, physical intelligence and size landings; the 100-year industry Agent innovation practices cover applications such as customer service, marketing, trading and enterprise production.
AIGC Content Industrialization Focus on Content Production, Quality Assessment, Security Compliance and MultiAgent Synergies; Financial AI Native Engineering Practice: from R & D to Business Breaking, the perspective extends to insurance, securities, cross-border payments and financial controls.
Technological change is changing the role of the organization and people.
In addition to the technical topics, the Congress set up the next stop for the development of the AI Native organization and technologist to discuss how enterprises could restructure collaborative mechanisms, engineering management and talent, and how technicians could move from code implementers to system designers and task managers.
In addition, the Congress established a special event such as the KIRO "Ghost Initiative" auction, AI's landing on the ground — the 9th anniversary of QCon's Shanghai Walking Night and the initiation of AI — to provide participants with opportunities for hands-on practice, real case discussions and face-to-face exchanges.
From architecture, infrastructure and data systems to R & D delivery, organizational change and industry applications, the current QCon Shanghai branch presented an AI Production Road Map: AI is gradually becoming a production system that can be designed, run, evaluated and continuously improved from a capability that needs to be mobilized.
Eco-mapping: exhibition space exchange and sponsorship support
On the floor of the General Assembly, the exhibition area is equally well populated for technical exchanges. Amazon Cloud Technology, TDengine (Tothic Data), Observation Clouds, Combining Information, Melting Technology, AIMA Community of IQDs, AreaL, IPIP, etc., develop platforms, data and algorithms and engineering tools around AI infrastructure, AI agent, demonstrate cutting-edge technology and practical programmes and interact in depth with participants. Partners such as Tencent clouds also participated in the Congress through sponsorship and ecological synergy.
In addition, the Central European Institute of International Business and Industry, EMBA, Licorne and TGO, as academic partners, supported the Congress. Multi-partners work together to link technological innovation and industry practices to promote the continued fall of AI technology in a real business landscape.
Thermal continuity: December AICon Beijing station
On December 18-19, the Aicon Global Congress on Artificial Intelligence Development and Applications will be held in Beijing. The Congress, which focused on exploring the A application boundary, set up 13 major technical topics covering the core directions of LLM post-training, reasoning optimization, Agent engineering, intellectual engineering, AI agent assessment and security, self-evolution systems, AI software delivery, native architecture and organizational change, bringing together first-line technical experts to share in depth the real practices of AI technological breakthroughs and enterprise size.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.