2026 年，AI 正在发生一个值得关注的变化。过去，我们更多讨论的是“模型能做什么”：代码能不能写、图片能不能生成、复杂问题能不能回答。而今天，当 AI 开始真正参与工作，软件工程面对的问题也随之发生变化，它面对的就不再只是一个模型调用问题，而是会进一步触及研发、数据、基础设施、质量、运维乃至组织协作的整个链条。这些也是今年 QCon 全球软件开发大会·2026（上海站）重点关注的方向。
10 月 22—24 日，来自阿里巴巴、腾讯、字节跳动、美团、小红书、蚂蚁集团、无问芯穹、阶跃星辰、快手、百度、网易、微软、华为、携程、平安等企业的一线技术实践者，将从 AI 应用、Agent 工程、AI Coding、基础设施、数据系统、质量、安全、SRE 以及具身智能等方向，分享 AI 进入真实生产环境之后正在发生的变化。
主题演讲
从 Coding 到消费：AI 的价值正在溢出
如果说 Coding 是这一轮 AI 最早被大规模验证的场景之一，那么接下来，AI 的影响，很可能延伸至更广泛的消费与业务场景。
阿里巴巴中国电商事业群飞猪 CTO 陈烨博士将分享《从 Coding 到消费：AI 的下一场革命》。本次演讲将从“可验证性”出发，分析 AI 从 Coding 走向 Agent，再进入消费场景的演进逻辑，消费级 AI 的难点究竟在哪里。
以旅行这样的复杂消费场景为例，AI 不仅需要理解用户意图，还需要连接供给与服务网络，并处理大量动态变化的消费信息。如何构建 AI-Native 的供给与服务网络？如何将静态数据与实时消费数据转化为训练资产？又如何通过后训练建立持续迭代的数据飞轮？这些问题最终指向的，并不只是一个更聪明的消费入口，而是 AI 如何真正进入交易、服务与履约。
当 AI 开始参与完整的消费链路，互联网产品的变化可能也不再只是增加一个 Copilot，而是重新组织人与服务连接的方式，并让真实消费场景成为 AI 持续进化的土壤。
从 Token 生产到生产力转化，基础设施如何创造更大价值
随着 Agent 应用加速爆发，Token 的使用场景正从单次推理扩展到持续推理、工具调用与复杂任务执行，词元工厂也面临着新的效率挑战。决定 AI 生产力的，不再只是模型本身的能力，更在于如何高效组织从能源、算力到 Token 产出，再到实际生产力转化的完整链路。
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，将带来《以 Agentic Infra 技术实践支撑词元工厂持续演进》的主题分享。他将从 Agent 时代的实际需求出发，基于无问芯穹系统化的技术实践，探讨 Agentic Infra 如何通过更高效的资源组织、推理优化与系统协同，持续提升从能源到 Token 的生产效率。并进一步通过底层技术创新推动 Token 的高效利用与价值转化，让 Agent 不仅能够生成内容，更能在复杂任务中持续优化、稳定运行并交付可复现的结果，推动 Token 更高效地转化为实际的生产力。
从工具到 Agent：内容创作正在经历怎样的变化？
当“生成视频”越来越容易，内容创作真正的难题是什么？小影科技创始人兼 CEO 韩晟将分享《从工具到 Agent：小影内容创作的三次范式迁移》。本次分享将以小影科技十余年的影像创作实践为主线，拆解从工具、生成到 Agent 的三次产品与技术迁移，并讨论当单点模型能力逐渐成为基础设施后，内容创作产品需要重新解决哪些工程问题：如何理解复杂创作需求，如何编排多种模型与工具，如何控制生成效果，如何利用用户数据形成反馈闭环，以及如何在真实业务中实现规模化交付。
分享将结合 AI 音乐 MV、移动场景化 Content Agent 等落地案例，介绍小影科技内部 AgentOps 在全球化市场运营稳定性、多 App 多场景快速迭代、端云协同与成本优化等方向的实践成果，并基于这些一线落地经验，拆解从 “AI 帮用户生成内容” 走向 “AI 协助用户完成创作” 过程中的探索与思考。
大会首日上午的三场主题演讲，恰好从不同方向勾勒出了这场变化的轮廓。
专题演讲
从“能不能做”到“能不能长期做好”：Agent 工程进入深水区
主题演讲之外，三天大会的大量实践都在回答一个问题：当 Agent 真正进入生产环境，除了 Prompt 和模型选择，还需要解决什么？
记忆、执行环境、工具调用、权限与安全、评测和可观测性，正在成为 Agent 工程的基础能力。更关键的是，Agent 的任务往往无法一次完成——执行可能出错，工具可能异常，线上环境也会持续变化。因此，生产级 Agent 需要从一次性的“输入—输出”，走向“规划—执行—观测—验证—反馈—再执行”的持续运行。
这也是今年大会频繁出现 Harness、Loop、Agent Evaluation、Agent Observability 等议题的原因。从阿里巴巴的生产级 Agent，到蚂蚁阿福 Harness 的线上 Badcase 处理，再到快手电商导购的 Harness Loop、网易智企 CodeWave 的 Loop Engineering，越来越多团队正在探索：如何让 Agent 从一次性任务执行，走向能够持续运行、持续反馈和持续改进的生产系统。
AI Coding 的下一阶段，不再只是“生成更多代码”
今年另一个非常集中的讨论方向，是 AI 对软件研发本身的重构。AI Coding 正在快速降低代码生产的成本，但代码生成越容易，验证、质量控制、系统设计和最终交付的重要性反而越突出。当“写代码”不再是研发流程中最稀缺的环节，真正的问题转向了“当 AI 大规模写代码之后，软件工程应该如何重新组织”。
今年上海站的 AI Coding 专题进一步进入大型工程的真实场景：小红书将分享大型客户端工程中的 AI Coding 实践；阿里巴巴将分享百万真实任务验证的 Open Code Review；蚂蚁集团将讨论 60 万行巨石应用的 AI 重构等等。
当 AI 成为新的计算主体：基础设施正在发生什么变化？
如果说 AI Coding 改变的是软件生产，那么 AI Infra 与数据系统改变的则是 AI 运行的底座。在第二天的议题中，可以看到非常完整的一条技术链：
- 从算子调优、推理优化、KV Cache、强化学习，到 Agent 执行环境；
- 从企业级 Agent Sandbox，到数据系统、Data Agent、Semantic Runtime；
- 从 Trace、评测，到生产流量中的实时观测
- ……
这些议题看似分散，实际上都在回答同一个问题：当 AI 成为新的计算主体，原有的基础设施还能不能直接承接？
数据也正在发生类似变化。过去，数据系统主要服务人和传统应用；而在 Agent 时代，AI 本身正在成为新的数据消费者、分析者和执行者。AI Native 可能最终并不只是“给原来的系统加一个 AI”。而是重新思考数据、计算、软件和人的关系。
真实世界的软件工程：边界还会继续扩展
AI 的边界也正在从数字世界向真实世界延伸。机器人、汽车、移动终端、工业设备，都开始成为 Agent 的运行环境。我们策划的专题方向也将围绕具身智能、端云协同和 AI Native 终端等展开，届时将有多场实践分享。
从大晓机器人、Dexmal 原力灵机的具身智能实践，到蔚来 NOMI 的端云协同，再到快手、蚂蚁、阶跃星辰等团队在终端与多端互联方向的探索，可以看到一个越来越清晰的趋势：
AI 正在从“理解数字世界”，走向“进入真实世界并对其产生影响”。而一旦 AI 开始面对真实环境，软件工程需要解决的问题也会进一步扩展——不确定的环境、实时反馈、设备协同、安全边界，以及更加复杂的系统状态。这可能是 Agent 之后，软件工程面对的又一个巨大变化。
软件工程的下一阶段，可能不是“被 AI 替代”，而是重新定义
回看今年上海站的议题，从 Agent、AI Coding，到 AI Infra、数据、SRE、质量、安全，再到具身智能与终端全面覆盖，如前文提到的，我们邀请了这些不同技术领域的一线实践者，把正在发生的技术变化带到现场。三天时间，欢迎来跟各企业/学术界的老师们面对面交流。带着你正想解决的问题来，带着真实的实践经验、技术方法和新的思考回去。
10 月 22—24 日，欢迎来到 QCon 全球软件开发大会·2026（上海站），期待现场见。
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。
In 2026, AI was in the midst of an interesting change. In the past, we talked more about what models can do: whether codes can be written, pictures can be generated, complex questions can be answered. And today, when AI begins to be truly involved, the problems facing software engineering change, and it is no longer just a question of modeling, but it will go further across the entire chain of research and development, data, infrastructure, quality, transport and even organizational collaboration. These are also the directions in which QCon Global Software Development Conference 2026 (Shanghai Station) focused its attention this year.
On October 22-24, front-line technology practitioners from Alibaba, Tencent, byte beat, group, REDnote, ant group, no-questioned core domes, step-up stars, quick-handers, Baidu, web-employer, Microsoft, Huacheng, transport, safety companies will share the changes that are taking place in the direction of AI applications, Agent engineering, AI Coding, infrastructure, data systems, quality, safety, SRE and physical intelligence.
Keynote address
From Coding to Consumption: The value of AI is spilling over
If Coding is one of the first large-scale validated scenarios for this round of AI, then the impact of AI is likely to extend to a wider consumption and business landscape.
Dr. Alibaba is sharing "The Next Revolution from Coding to Consumption: AI." This presentation will start with "validability" and analyze the evolutionary logic of AI moving from Coding to Agent and then into the consumer scene.
In the case of complex consumption scenarios such as travel, AI needs to understand not only user intentions, but also to connect to supply and service networks and to process large volumes of dynamically changing consumption information. How to build an AI-Native supply and service network? How can static data and real-time consumption data be translated into training assets? How do you train to create continuous and iterative data flyers? These questions ultimately point to not just a smarter consumer entry, but how AI actually gets into transactions, services and compliance.
When AI begins to participate in a complete consumption chain, changes in Internet products may no longer simply add a Copilot, but rather a way to reorganize people and services and to make real consumption scenarios the soil of continuous evolution of AI.
From Token production to productivity transformation, how infrastructure creates greater value
With the acceleration of the Agent application, Token ' s use scene is expanding from a single reasoning to ongoing reasoning, tool call and complex missions, and the meta-factor is facing new efficiency challenges. AI productivity is determined not only by the capabilities of the model itself, but also by how to organize efficiently from energy, computing to Token output to the complete chain of real productivity transformation.
Associate Professor of the College of Artificial Intelligence of Shanghai University of Transport, Co-founder and Chief Scientist of the Unquestioned Coding Dome, Dai Kuo Ho, will lead to the theme-sharing of the “Continuing evolution of the meta-factory factory with the Agenic Infra Technical Practice”. From the real needs of the Age of Agent, he will explore how the Agenic Infra can continuously improve production efficiency, from energy to Token, through more efficient resource organization, reasoning and system synergy, based on the technical practice of the no-question core dome system. Token ' s efficient use and value transformation is further promoted through bottom-up technological innovation, enabling Agent not only to generate content, but also to continuously optimize, stabilize and deliver recoverable results in complex missions, and to move Token more efficiently into real productivity.
From tool to agent: how is content creation going through?
What is the real problem of content creation when it becomes easier to generate video? The founder and CEO of small film technology will share "From Tools to Age: Three Model Migrations of Small Film Content." This sharing will be based on a 10-year-old video creation exercise in small-film technology to dismantle three movements of products and technologies from tools, generation to Agent, and to discuss what kind of engineering needs to be addressed when single-point model capabilities become infrastructure: How to understand complex creative needs, how to structure multiple models and tools, how to control the production effect, how to use user data to create feedback loops, and how to achieve scale delivery in real business.
The sharing of examples of landings that will be combined with AI music MV, mobile landscapes, etc., presents the results of the work done by Agent Ops within small film technology in the direction of stability in the globalized market, multiple App multi-species, cloud synergism and cost optimization, and deciphering the search and reflection from "AI help users generate content" to "AI help users to create."
The three keynote addresses on the first morning of the Conference coincided with the outline of the change from different directions.
Keynote address
From "Can it be done" to "Can it be done in the long term": Agent works into deep water
In addition to the keynote address, a large body of General Assembly practice over the past three days is answering the question: When Agent actually enters the production environment, what needs to be done beyond Prompt and model selection?
Memory, implementation environment, tool call, permission and safety, assessment and observation are becoming the foundational capabilities of the Agent project. More crucially, Agent's tasks are often not completed at once — implementation can be wrong, tools can be unusual and the online environment can change continuously. Therefore, the production level Agent needs to move from a one-time “input-out” to a “planning-implementation-observation-test-feed-back-implementation”.
This is also the reason for the frequent occurrence this year of topics such as Harness, Loop, Agent Evaluation and Agent Observacity. From Alibaba production grade Agent to the ants Alfred Harness online processing of Badcase, to Halness Loop, a fast-track electrician, Loop Engineering of CodeWave, an increasing number of teams are exploring how to move Agent from a one-time mission to a production system that can be continuously operated, continuously fed and continuously improved.
The next phase of AI Coding is no longer just " Generate More Codes"
Another very focused discussion this year is the re-engineering of the software itself. AI Coding is rapidly reducing the cost of code production, but the easier the code is generated, the greater the importance of validation, quality control, system design and final delivery. When “writing code” is no longer the most scarce link in the R & D process, the real question turns to “how software works should be reorganised when AI writes code on a large scale”.
This year, the subject of Al Coding at Shanghai Station went further into the real scene of large-scale engineering: REDnote would share the AI Coding practice in large-scale client projects; Alibaba would share the Open Code Review of millions of real task; and the ants group would discuss the AI re-engineering of 600,000 giant stone applications.
When AI becomes the new subject of calculations: What is going on in infrastructure?
If AI Coding changes software production, AI Infra and the data system changes the base of AI running. On the topic of the next day, a very complete technological chain can be seen:
- From arithmetical optimization, reasoning optimization, KV Cache, intensive learning, to the Agent implementation environment;
- From Enterprise Sandbox to Data Systems, Data Agent, Semantic Runtime;
- from Trace, evaluation, to real-time observations in production flows
- ……
These issues appear to be scattered and are in fact answering the same question: when AI becomes the new subject of calculations, can the original infrastructure be taken over directly?
Similar changes are taking place. In the past, data systems served primarily as service providers and traditional applications; in the Age of Agent, AI itself was becoming a new data consumer, analyst and implementer. AI Native may end up not just adding an AI to the original system. It's about rethinking data, computing, software and human relationships.
Software engineering in the real world: the boundaries will continue to expand.
The boundaries of AI are also extending from the digital world to the real world. Robots, cars, mobile terminals, and industrial equipment are beginning to become Agent operating environments. The thematic orientations that we plan will also revolve around intelligence, cloud synergetics and the AI Native Terminal, where there will be multiple practice sharing.
From the smart practices of the Big Mixer robots, the Dexmal Force Spirit, to the confluence of wind clouds at the end of the NOMI, to the search of fast hands, ants, leap stars, etc. in the direction of end-to-end interconnection, an increasingly clear trend can be observed:
AI is moving from “understanding the digital world” to “entering and influencing the real world”. Once AI begins to face the real environment, the issues that software engineering needs to address will also expand — uncertain environments, real-time feedback, equipment synergy, secure boundaries, and more complex systems. This could be another huge change in software engineering after Agent.
The next phase of the software project may not be "altered by AI" but redefinition
Looking back at the issue of Shanghai Station this year, from Agent, AI Coding, to AI Infra, to data, SRE, quality, security, to full coverage with intelligence and terminals, we invited, as mentioned earlier, front-line practitioners in these different technical fields to take the technological changes that are taking place. For three days, we would welcome a face-to-face conversation with teachers from businesses/academics. Take with you the problems you are trying to solve and go back with real practical experience, technical methods and new thinking.
Welcome to the QCon Global Software Development Conference, 2026 (Shanghai Station), 22-24 October, and look forward to seeing them on site.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.