InfoQ AI、ML 与数据工程趋势报告 - 2026

InfoQ AI/ML 年度趋势报告为 InfoQ 读者全面概述了 AI、ML 和数据工程领域的新兴趋势与技术。本报告总结了 InfoQ 编辑团队与外部嘉宾进行的播客对谈，探讨 2026 年 AI 和 ML 技术的发展趋势，以及未来十二个月值得关注的内容。与报告和趋势图配套的播客还深入讨论了所有这些趋势。The InfoQ AI ML Annual Trends Reports offer InfoQ readers a comprehensive overview of emerging trends and technologies in the areas of AI, ML, and Data Engineering. This report summarizes the InfoQ editorial team’s podcast with external guests to discuss the trends in AI and ML technologies for the year 2026 and what to look out for in the next twelve months. In conjunction with the report and trends graph, our accompanying podcast features insightful discussions of all of these trends.

AI 和 ML 趋势采用情况图表AI and ML Trends Adoption Graph

年度趋势报告的一个重要组成部分是趋势图，它展示了进入创新者类别的趋势、技术和主题，以及晋升至早期采用者和早期大众类别的项目。这些类别基于 Geoffrey Moore 的著作《跨越鸿沟》。InfoQ 主要关注尚未跨越鸿沟的类别。以下是今年的趋势图：An important part of the annual trends report is the trends graph, which shows the trends, technologies and topics that have made it to the innovators category and which have been promoted to the early adopters and early majority categories. The categories are based on the book, Crossing the Chasm, by Geoffrey Moore. At InfoQ, we mostly focus on categories that have not yet crossed the chasm. Here is this year’s graph:

自 2022 年 11 月 ChatGPT 发布以来，差不多已经过去四年。RAG 技术开始在企业中得到应用、自主 LLM 智能体问世，也已是两三年前的事了。过去几年里，AI 的爆发式增长在 2026 年仍在持续。自 InfoQ 团队去年讨论趋势报告以来，AI 技术也取得了重大创新与发展（见下方 2025 年趋势图）。It has been almost four years since ChatGPT came out back in November 2022. It has also been two or three years since RAG Technologies started seeing enterprise adoption and autonomous LLM agents came out. The explosive growth of the last few years has continued in 2026 and AI technologies have seen significant innovations and developments since the InfoQ team discussed the trends report last year (see 2025 trends graph below).

有关 2025 年 AI/ML 技术趋势的详情，也请参阅去年的趋势报告文章。For the details of 2025 AI ML technology trends, also check out last year’s trends report article.

今年的趋势文章在去年讨论的内容基础上进行了扩展。文章重点介绍了趋势图，该图展示了技术采用的不同阶段，并进一步详述了自去年趋势报告发布以来新增或更新的各项技术。我们还讨论了哪些技术和趋势在采用图中得到了提升。在今年更新的采用图中，团队还添加了一些我们在播客中没有讨论的主题。This year’s trends article builds on what we discussed last year. It highlights the trends graph, which shows different phases of technology adoption and provides more details on individual technologies added or updated since last year’s trends report. We also discuss which technologies and trends have been promoted in the adoption graph. For this year’s updated adoption graph, the team has also added a few additional topics that we didn’t discuss in the podcast.

你会注意到，我们还对范围有所重叠的主题进行了大幅重组，并移除了一些技术。由于这些技术已被广泛采用，因此不再需要在采用情况图表中跟踪它们。As you will notice, we have also done some major reorganizing of the topics where there is an overlap of the scope, and removed a few technologies that no longer need to be tracked on the adoption graph because of their mainstream adoption.

以下是自去年报告以来的一些变化亮点：Here are some highlights of what has changed since last year’s report:

创新者Innovators

在我们的采用曲线中，第一个类别是创新者，今年新增了几个有趣的话题。下文将详细讨论这些话题。The first category on our adoption graph, Innovators, has a few interesting new topics added this year. These topics are discussed in detail below.

智能体间协议Agent-to-Agent Protocols

随着智能体式人工智能系统日趋复杂，涉及多个不同的智能体和领域，单个智能体之间以及多智能体系统之间的通信，对于确保这些系统始终安全、准确、可靠至关重要。As agentic AI systems become complex involving multiple different agents and domains, the communication between the individual agents as well as between multi-agent systems is crucial for these systems to be consistently safe, accurate, and reliable.

我们趋势报告播客的一位嘉宾 Daniel Dominguez 谈到了智能体通信协议，以及它们对整个智能体 AI 解决方案领域的重要意义。One of our panelists in the trends report podcast, Daniel Dominguez, spoke about agent communication protocols and the significance they have in the overall agentic AI solutions.

Daniel Dominguez：在这一切之下，架构正在发生转变。要让智能体在企业内部发挥作用，它们需要访问工具、数据和 API，并与其他智能体交流。MCP 或 A2A 等协议都将变得越来越多样化。我想，关键问题在于，智能体协议是否会成为一种新的企业集成层。比如，今天我们将应用连接到 API；明天，我们或许会进一步发展，在这些 API 之上加入智能体发现能力，用于协调任务。我认为，所有这些架构层也正在改变智能体彼此交互的方式。

Daniel Dominguez: There is an architectural shift happening underneath all of this. For agents to become useful inside enterprises, they need to access tools, data, APIs, and talk to other agents. All the protocols like MCP or A2A are going to become more and more different. I think the question would be whether agent protocols become a new enterprise integration layer.

For example, today we connect applications to APIs. Tomorrow we may increase and have agent discovery capabilities that we see in task coordinating work on top of those APIs. I think all of these architectural layers are also changing the way that agents interact with each other.

推理引擎（vLLM、SGLang、llm-d）Inference Engines (vLLM, SGLang, llm-d)

另一个取得显著进展并发生转型的领域是 AI 推理。2026 年被称为推理之年。各家公司正从训练模型转向在生产环境中运行 AI 推理。预计到今年年底，推理将占 AI 计算支出的约三分之二。推理模型、长上下文和智能体已将推理从模型执行问题转变为一个涵盖计算、内存、网络、路由和经济效益的系统问题。Another area that saw a significant development and transformation is the AI Inference. 2026 has been called the year of inference. The companies are shifting from training models to running AI inference in production. Inference is expected to account for around two-thirds of AI compute spending by the end of this year. Reasoning models, long context, and agents have turned it from a model-execution problem into a systems problem spanning compute, memory, networking, routing, and economics.

Hien Luu详细讨论了这一趋势，并分享了他在推理框架方面观察到的情况。Hien Luu discussed this trend in detail in what he is seeing in terms of inference frameworks.

Hien Luu：自 2026 年初以来，推理领域迅速升温。显然，一方面，vLLM、SGLang 等各种推理技术栈，正在帮助我们这些人搭建自己的推理服务，并部署在自己的环境中。与此同时，这些开放权重模型的能力也越来越强。起初，了解这类技术如何运作、如何部署的人并不多，但现在网上有了更多相关信息和资源，介绍如何实际使用这些软件，以及如何对它们进行优化。 与此同时，我们讨论的一些应用场景涉及金融公司等持有个人身份信息（PII）的机构。对于这类特定场景而言，在内部部署开放权重模型非常切合需求。我相信，随着这件事变得更容易接触、也更容易实现，企业会继续增加这类应用，这种趋势也会持续下去。有些公司出于节省成本的考虑，实际上希望在内部部署开放权重模型，例如用于编程。在我看来，出于种种原因，推理将继续成为一个非常实用、非常强大的领域，未来会有越来越多的公司开始这么做。

Hien Luu: The inference topic has exploded since the beginning of 2026. Obviously from one end, you have all these inference stacks from vLLM, SGLang, and so on that are helping the rest of us to spin up our own inference service in our own environment. At the same time, these open-weight models are becoming more and more capable. That space is initially not a whole lot of people how it works and how to operate it, but now there's a lot more information and resources out there about how to actually use these software, [and] how to optimize them. At the same time, some of the use cases we talk about are the financial companies where you have PIIs.

The ability to host these open-weight models internally for those kinds of specific use cases are very, very relevant. I believe that that's going to continue to grow, continue to be the case where a company will do more of this moving forward now that it's a lot more approachable, easier to do. There are some companies that because they want to save cost, they actually want to host internal open-weights models for coding, for example. For many reasons that inference is going to continue to be a very, very useful, very powerful topic and more and more companies will start doing this moving forward in my mind.

测试框架工程Harness Engineering

一个相对较新的领域是 Harness Engineering，自今年以来一直备受关注。Hien 谈到了这一主题的重要性，以及它如何支持内存管理和协同执行等工程能力。A relatively new area that has been seeing a lot of attention starting this year is Harness Engineering. Hien talked about the importance of this topic and how it supports the engineering capabilities like memory management and co-execution.

Hien Luu：我想在关于智能体的讨论中补充[两点]。其中一点是 harness engineering。我认为它是对上下文工程的延伸。不过我觉得有件事很有意思，也和我即将举办的一场研讨会相关：这些智能体 SDK 起初只是帮助我们编写循环逻辑。如今，这些 SDK 已经扩展到提供 harness engineering 功能，包括上下文的内存管理、协同执行，以及一些开箱即用的工具，用于直接操作本地文件系统。你还可以看到另一种趋势：LangChain、Anthropic 和 OpenAI 的这些智能体 SDK，都在朝这个方向发展，把更多能力纳入 harness engineering。今后任何构建智能体的人都应该留意并加以利用。另一个完整的趋势是如何管理这些智能体。这些公司正在提供运行环境，让你托管自己的智能体。你是否想这么做是另一个问题，但这是他们正在关注的方向：“我们能够构建，也能够把它构建好；接下来该如何运营，并让它大规模运行？”

Hien Luu: I want to add [two things] to discussions of agents. One is harness engineering. I think that extensions of context engineering. But the thing that I see is quite interesting and relevant to a workshop that I'm going to be giving is that these agent SDKs start out with just helping us to write the loop. Now this SDK has been expanded to provide harness engineering capabilities, including memory management for context, including co-execution, including some of the out-of-the-box tools for doing into actual local file systems.

That’s another trend you can see that these agent SDK from LangChain or Anthropic or OpenAI; they all move in that direction of adding more capabilities as part of the harness engineering. As anyone building agents moving forward, that's something to look out [for] and leverage. There’s a whole other trend of managing these agents. These companies are providing a runtime environment for you to host your agent. Whether you want to do it or not, that's a separate question, but that's what they're seeing. "We are able to build, we are able to build it well, and then how do we operate and run it at scale?"

规范驱动开发与 AI 软件工厂Spec-Driven Development and AI Software Factories

随着 AI 技术在软件开发过程中承担了大部分编码工作，我们看到开发者开始更加关注需求和规格说明，以充分利用自动化技术完成应用中的设计和编码任务。With AI technologies contributing to most of the coding in the software development process, we are seeing the developers starting to focus more on the requirements and specifications to take full advantage of automating the design and coding tasks in their applications.

这一技术转变催生了规范驱动开发和 AI 软件工厂等趋势。一些组织正在组建专门团队，以打造软件工厂模式，使开发团队能够使用 AI 工具自动完成整个软件开发流程。该模式涵盖 SDLC 中的具体步骤，例如自动梳理并分析现有软件应用的业务领域和需求，审查新的业务需求或系统增强需求，为这些新需求编写功能和非功能（技术）规范，生成用于实现新需求的设计和开发任务，最后自动生成更新后的软件应用代码及其他制品。This technology shift has given rise to trends like Spec-Driven Development and AI Software Factories. Some organizations are creating dedicated teams to create the software factory model where the development teams can use AI tools to automate the complete software development process. This model includes the specific SDLC steps like automatically compiling and analyzing the business domain and requirements of an existing software application, reviewing the new business requirements or systems enhancements, writing functional and non-functional (technical) specifications for those new requirements, generating the design and development tasks to implement the new requirements, and, finally, auto-generating the code and other artifacts of the updated software application.

嘉宾 Hien Luu 分享了他对规范驱动开发的见解。Panelist Hien Luu shared his insights about Spec-Driven Development.

Hien Luu：这大概是四五年前开始的。如今的编程代理已经相当强大。不过，在构建项目时，通常你会想提前规划好自己想要什么、需要哪些功能等等，而不是马上就让编程代理开始构建。这其实是把我们所熟知的良好实践借鉴过来，并将其整理成你所使用的编程代理能够理解的形式。举例来说，如果你要从零开始构建一个全新的项目，可以先写一份规范：项目是做什么的？你想要哪些功能？然后和编程代理进行一次头脑风暴，为它提供所需的背景信息，再真正开始构建。我认为人们围绕这一做法总结出了很多很好的实践。大多数人都认为这是良好的实践。我现在不太常听到更多相关讨论，因为大家正在实际运用它；而 AI 工厂则是近几个月兴起的另一种做法，旨在探索如何将 AI 融入软件开发生命周期的各个环节。典型的用例是：“嘿，有人报告了一个 bug。如何让 AI 从头到尾处理这件事，接手这个 PR、进行分析、尝试复现问题、提出修复方案并提交 PR，然后你再继续完成评审、测试等后续步骤？”我们都有太多事情要做、太多任务要处理。我认为，如果能帮助自动化这些工作，就能从另一个维度加速软件开发生命周期。

Hien Luu: This probably started maybe four or five years ago. The coding agent now is pretty powerful. But as you build a project, typically you want to have a plan in advance in terms of what you want, the capabilities and whatnot, and not jump to the coding agent immediately to build. It is borrowing what we know as a good practice to do, but now codified it in a form for the coding agent that you work with. If you build a brand new greenfield project, for example, you build a spec up, what is it about? What are the capabilities that you want?

Then you have a brainstorming session with your coding agent to provide a context that it needs before you actually go and build it. I think there's a lot of good practices there that people have built around this. I think most people recognize that's a good practice. I don't hear much more about it now because as people are actually practicing it and the AI factory is another that's probably in the last few months of really how to infuse AI into the various steps of your software development life cycle.

The typical use case is, "Hey, somebody raised a bug. How do you get AI to take that from beginning to end to actually take the PR, do an analysis, try to recreate the problem, come up with a fix, submit a PR, and then you go through the rest of the steps of review and testing and so on?" That is as we all have too many things to do, too many tasks. I think if we can help automate that, that will accelerate another dimension to accelerating the software development life cycle.

物理人工智能 / 具身人工智能Physical AI / Embodied AI

另一个持续涌现越来越多创新的领域是物理 AI。我们要在这里将智能带入现实世界。AI agent 是一种软件系统，那么我们该如何将同样的能力带入硬件中，无论是机器人、协作机器人（cobot），还是自动驾驶汽车？Another area that continues to see more and more innovation is Physical AI. This is where we bring intelligence into the real world. An AI agent is a software system, how do we bring the same power into the hardware, whether they're robots or collaborative robots (cobots) or self-driving cars.

Daniel Dominguez 谈到了物理 AI 领域正在发生的事情，以及我们应该关注的重点。Daniel Dominguez talked about what's happening and what we should be paying attention to in the Physical AI world.

丹尼尔·多明格斯：基本上，它是在问：“我们如何将 AI 带入物理世界？”过去几年里，我们一直致力于将智能融入软件。现在，我们开始把这种智能连接到能够感知物理世界并在其中采取行动的机器上。你刚才提到，我们开始看到机器人、自动驾驶汽车、无人机和工业系统都加入了这一转变。我认为，物理 AI 与软件智能体有着根本区别：物理 AI 出错会造成实际的物理后果。在软件领域，智能体可能会在生成自己的文件时出错；但机器人可能会驱动自己的执行器。这会是完全不同级别的错误。因此，可靠性、仿真、安全、验证和评估，在物理世界中会变得越来越重要。 基于这一点，目前正在研发的各类机器人基础模型，还有一个重要的转变：机器人技术开始受益于与推动语言和视觉领域变革相似的基础模型方法。我们将开始看到诸如用于机器人的数字孪生仿真、开放机器人领域将发生的种种变化，以及从 AI 智能体走向物理智能体将带来的种种发展。未来一年，这一切都会开始以越来越快的速度推进，因为 AI 已经来到我们身边。现在，我们必须把它带入物理世界。

Daniel Dominguez: Basically it's saying, "How do we bring AI into the physical world?" Over the last few years we have spent the time putting intelligence into software. Now we're starting to connect that intelligence to machines that can perceive and act in a physical world. So what, you said, we are starting to see robotics, we're starting to see autonomous vehicles, drones, and industrial systems are now part of this transition. What I think makes physical AI fundamentally different from software agents is that mistakes have physical consequences on the physical AI.

One thing is that in software, an agent can make a mistake generating their own file, but a robot can move their own equator. It’s going to be a whole different level of mistakes. Reliability, simulations, safety, verification, [and] evaluations are going to become more and more important in the physical world. Based on that, all the foundation models for robots that are happening right now, another important shift is that robotics are starting to benefit from more of the same foundational model approach that transformed the language and vision.

We're going to start seeing things like [a] simulation for digital twins robots, all the things that are going to happen for open robotics, [and] all the things that are going to happen from AI agents to physical agents. All of this is just going to start moving faster and faster over the next year because right now the AI is here. Now we have to put it in our physical world.

Chip Huyen 还讨论了机器人和视觉语言模型如何协同工作。Chip Huyen also discussed how robots and vision language models can work together.

Chip Huyen：让模型进入现实世界是下一个前沿领域，我们也看到了很多进展。我认为唯一的问题是，挑战有好几个。我们总说机器人就像汽车，或者说就像 LLM。我们会说：“好吧，我们知道现实世界里有很多机器人。”到处都有 Waymo，也有 Zoox。我们在现实世界里看到了很多机器人。机器人模型和 LLM 之间有一个很大的区别，那就是回溯的难题。对于 LLM 来说，如果你在虚拟环境，也就是电脑上训练模型时，它犯了错，你可以回退。你可以重置状态，然后继续。你可以进行大量探索，挑选出最佳状态。但在机器人领域，你没法完全做到这一点。你不能对机器人说：“好吧，那里有一段楼梯。”机器人摔下去就坏了。你不能简单地把它倒回去，让它恢复成之前的样子。这让探索变得极其困难。另一个挑战是数据。我一直在研究很多机器人系统。我认为 Janatics Robotics 公司——我们还有一批公司，比如 Generalist、Physical Intelligence、Sunday 和 Skild AI，我对它们都非常期待。

Chip Huyen: It's the next frontier to get the models into the real world and we see a lot of progress. I think the only challenge is that there are several challenges. We keep saying that robotics are like cars, like LLM to say, "Okay, we know we have a lot of robots in the real world." We have Waymos everywhere. We have Zoox. We see a lot of robots in the real world.

One big difference between robotics, models, and LLMs, it's that the challenge of backtracking. For LLMs, when you're training on a virtual, on the computer, if it makes a mistake, you can run back. You can reset the state and continue. You can do a lot of explorations and pick the best state.

You can't really do [that] exactly [in] robotics. You can't get a robot, "Okay, there's a staircase that's there". The robot falls and it breaks. You can't just run it back and get the same robot as before. It makes exploration extremely, extremely hard. Another challenge is data. I have been looking into a lot of robotic setups. I think Janatics Robotics company – we have a bunch of companies like Generalist, Physical Intelligence, Sunday, Skild AI, which I'm very excited about.

我们今年在创新者类别中新增或修订的其他趋势包括以下主题：Other trends in the Innovators category we have added or revised for this year include the following topics:

• 智能体安全与治理
  Agent Security and Governance
• 语音代理
  Voice Agents
• 专用决策模型（例如 TypeSafeAI 的 Jev 模型）
  Specialized Decision Models (e.g., TypeSafeAI's Jev Model)
• AI 网关
  AI Gateways
• GraphRAG
  GraphRAG
• 大语言模型运维 / 生成式 AI 运维
  LLMOps / GenAIOps
• 脑机接口
  Brain-Computer Interfaces

早期采用者Early Adopters

在早期采用者类别中，我们希望重点介绍两个主要主题：语言模型创新和 RAG。In the Early Adopters category, we would like to highlight two main topics: Innovations in language models and RAG.

AI 智能体 / 智能体编排AI Agents / Agent Orchestration

与去年类似，智能体 AI 今年再次成为主导话题。2025 年及此前几年，人们关注的都是如何学习和构建 AI 智能体。如今，各组织正试图弄清楚如何运维这些智能体、如何将它们部署到生产环境中，以及如何在生产环境中扩展它们。Similar to last year, Agentic AI has again been the dominant topic this year. 2025 and prior years were all about how to learn and how to build AI agents. Now the organizations are trying to figure out how to operate the agents, how to deploy them to production, and how to scale them in production.

Daniel Dominguez 在今年的播客中讨论了 AI Agent 领域的最新进展：Daniel Dominguez discussed the recent developments in AI Agents area, in this year's podcast:

Daniel Dominguez：我认为，2026 年最大的变化是，AI 正从我们与之交互的事物，变成我们可以把任务委派给它的事物。正如 Hien 所说，我认为过去几年大多数 AI 应用都建立在简单交互的基础上：你只要问一个问题，手机就会给你一个答案。而从今年开始，我们已经开始看到这样一种系统：它能够制定计划、使用工具、执行操作、独立完成工作，并持续工作更长时间。我认为，今年的重点不再是模型智能又有了多大的飞跃，而是如何把这种智能转化为真正可用的系统。我觉得有意思的是，智能体正在迅速跨越多个层级。最初是帮助个人的个人智能体，随后我们开始看到智能体与团队协作，而现在，各家公司正在考虑让智能体在整个组织范围内运行。我认为，重要的变化在于，我们正在超越单纯的聊天机器人模型。我们不再只是生成一个答案。智能体可以设定目标、使用工具、与软件交互、做出决策，并持续工作，直到任务完成。我认为，这让 AI 从一种界面转变为更接近执行层的东西。软件开发领域最能体现这一点。软件开发大概是这一转变最明显的领域。我们从自动补全发展到副驾驶，现在又有了编程智能体。Claude、Codex、Cursor、Replit 等工具能够浏览代码仓库、修改多个文件、运行命令、执行任务、检查故障，并自行迭代。如今有意思的问题已经不再是“AI 能写代码吗？”我们早就知道 AI 能写代码。在 AI 智能体领域，由多个智能体组成的、更复杂的系统（多智能体系统）正变得越来越常见，因为与基于单个智能体的系统相比，它们能带来更多优势。

Daniel Dominguez: I think the biggest change in 2026 is that AI is moving from being something that we interact [with] to something that we delegate [to]. Like Hien was saying, I think most AI applications over the last few years were based on a simple interaction. You will just ask a question and the mobile gives you an answer. Right now we're starting this year, we are starting to see a system that can plan, can use tools, can execute actions, can do their own work and continue working for a much longer period of time. I think this year is less about another jump in the model intelligence and more about turning that intelligence into actual systems.

What I find interesting is that agents are moving through several levels very quickly. We started with personal agents helping one person, then we started seeing agents working with teams, and now companies are thinking about agents operating across entire organizations. I think the important change is that we're moving from beyond just a chatbot model. Instead, we are simply generating an answer. An agent can have goals, can use tools, can interact with software, can make decisions, [and] can keep working until the task is complete.

I think that changes AI from [an] interface into something closer to an execution layer. We can see this mainly in software development. Software development is probably where this transition is more visible. We moved from autocomplete to copilots and now we have coding agents. We have tools like Claude, Codex, Cursor, Replit, and others that can navigate repositories, can modify multiple files, can run comments, can execute tasks, can inspect failures, and can iterate on their own. The interesting question is no longer, can AI write code? We already know that AI can write code.

In the AI Agents space, more complex systems using multiple agents (multi-agent systems) are becoming more common because of the additional benefits they can bring to the table, compared to systems based on individual agents.

小组成员 Arthur Casals 谈到了多智能体系统领域当前正在开展的研究、这些系统的工作原理，以及如何在现实世界中运用它们。Panelist Arthur Casals spoke about the current research happening with multi-agent systems, how these multi-agent systems work, and what are the real-world use cases to leverage them.

Arthur Casals：谈到多智能体系统这个主题，它在 LLM 智能体出现之前就已经存在了几十年。学术界的人看到 LLM 智能体以及 Transformer 出现后带来的各种东西，以及与整个多智能体系统领域相关的内容，便试图重新发明过去已有的东西。我的研究领域是经典多智能体系统，也就是 LLM 出现之前的多智能体系统；但这并不意味着它不能纳入或使用 LLM 系统作为其中的一部分。这些天来，我一直在研究信念修正。这实际上是解决 AI 领域两个问题的重要部分：一个与幻觉有关，另一个与信息经过智能体处理后的准确性和溯源有关。这一领域介于计算机科学和哲学之间，涉及大量形式逻辑，不过确实非常有趣。我认为，从今年年初到去年年底，我们一直在研究的一件事，是一种基于形式逻辑的档案定义，可以供人使用。它类似于我们之前讨论过的 Jev 相关概念。它有结构化输入和结构化输出，而我们要探讨的是，如何让 AI（而不是智能体）避免使用智能体，而是从这里开始，一路处理到那里。

Arthur Casals: When you think about multi-agent systems as a topic, it's been there for a few decades already before LLM agents. People from the academic community are seeing whatever comes with LLM agents and after transformers and is related to the entire multi-agent systems trying to reinvent things that were there in the past. My research area is on classical multi-agent systems, before LLM, which doesn't mean that it can't incorporate or use LLM systems as part of it. But these days I've been working on belief revision. This is a big part of solving two problems actually in the AI landscape, one related to hallucination and the second one related to information accuracy and provenance after being processed by the agent. It's between computer science and philosophy, a lot of formal logic, but yeah, it's really interesting. One thing that we've been working with, I think since the beginning of this year, and the end of last year, is a formal logic-based profile definition that could be used. It's similar to the concepts around Jev that we were discussing before. So [they] have a structure input, a structural output and how we can make the AI not the agents refrain from using agents, but actually working from here to there.

模型与智能体的 AI 评估AI Evaluations for Models and Agents

今年，我们也看到了质量保证领域的更多进展，包括 AI 评估、护栏机制和测试框架。2026 年被称为 AI 评估之年。该领域有许多不同的技术正受到关注。测试框架大概是今年讨论最多的话题。随着 AI 工具越来越多地用于编写代码和软件解决方案的其他组件，对正式、严谨、一致且可靠的验证框架的需求变得前所未有地迫切。那么，我们该如何衡量 AI 的有效性？在具备适当检查与制衡机制的端到端流程中，哪些指标有助于评估和验证？This year, we have also seen more developments in the quality assurance space with the AI evaluations, guardrails and harness frameworks. The year 2026 has been dubbed as the year of AI Evaluations. There are a lot of different techniques that are gaining attention in this area. Testing harness frameworks are probably the most talked about topic this year. As AI tools are increasingly leveraged to write code and other components of software solutions, the need for formal, rigorous, consistent and reliable validation frameworks has never been more critical. How do we then measure AI effectiveness? What kind of metrics can help in the evaluation and validation steps with an end-to-end process that has proper checks and balances?

我们请小组成员谈谈他们对智能体 AI 评估领域现状的看法。以下是他们的观察。We asked the panelists to explain what they are seeing in the agentic AI evaluation space. Here are their observations.

Hien Luu：对于那些正将 AI 智能体投入生产环境的公司来说，这是一个非常重要的话题，尤其是面向客户的智能体，而不是内部使用的智能体。以 DoorDash 为例，他们开发了一款用于订餐的 AI 助手。还有其他公司在为外部用户开发客户支持智能体。我们都知道，LLM 并非确定性的；在构建智能体时，它们可能会走错方向，也可能会错误地调用工具，诸如此类。评估可以帮助我们确保智能体在执行相应任务时能正确行事，并尽可能达到最高水平。没有评估，你怎么知道它是否偏离了预期，或做出了意料之外、违反公司政策之类的事情？公司不仅必须开展评估，还必须将其纳入构建 AI 智能体的核心流程。许多公司最终会构建一种称为评估框架的工具，用它来推动整个评估流程：从获取数据、运行评估或测试（我们都知道测试是什么，对吧？），到最后评估质量。

Hien Luu: This is a very important topic for companies that are bringing their AI agents to production, particularly the one that for their customers, not so much about internal ones. If you look at an example from DoorDash, they build an AI assistant for ordering food. You have other companies that are building customer support agents for external users. We all know that LLMs are not deterministics and building agents, they can take the wrong turn, they can make the wrong tool calls and stuff like that. Evaluations are a way to help us to make sure that the agents do the right thing for the right tasks at the highest level.

Without evaluation, how would you know that it's gone rogue or does something that's unexpected, not according to your company's policies and all that stuff like that. Companies not only have to do this, but they have to incorporate this as one of their core steps in building AI agents. A lot of companies end up building something called an evaluation harness, which helps drive the whole evaluation flow from getting the data, running the evaluation or tests, as we all know tests – right? – and then assessing the quality at the end.

Daniel Dominguez：我完全赞同这样的看法：今年正在成为模型和智能体的评估之年。一旦智能体能够执行操作，只评估模型就不够了。你必须评估整个系统，包括模型、工具、上下文、智能体循环以及环境。指标也因此发生了变化。我们关注的不再只是答案是否正确，还要看智能体是否完成了任务、花费了多少、耗时多久、需要多少次干预，以及过程中是否做出了不安全的操作。我认为，我们才刚刚开始探索如何评估所有这些与智能体、大语言模型以及它们在工作中进行的交互相关的工作流程。

Daniel Dominguez: I completely agree with the idea that this year is becoming the year of evaluations for models and agents. Once an agent can perform an action, evaluating only the model is not enough. You have to evaluate the whole system, you have to evaluate the model, the tool, the context, the agent loop, the environment. Here is where the metrics changed as well.

It’s not only whether the answer was correct, it's whether the agent completed the task, how much it cost, how long it took, how many interventions were necessary, and whether it did something unsafe along the way. I think all these evaluations were just starting to see how we are going to evaluate all this workflow regarding agents and regarding LLMs and regarding the interactions that they do in our job.

除了这些重点关注的技术趋势之外，还有几项技术正从创新者类别转入早期采用者类别。以下列出的所有技术都已历经广泛开发并得到普及，如今正被各类组织中的众多应用所采用。Outside of these highlighted technology trends, there are a few entries that we are moving from Innovators into the Early Adopters category. All these technologies listed below have gone through extensive development and adoption and are now being used by many applications in various organizations.

人工智能技术已开始应用于生产支持、故障排查、维护和升级运维中的 DevOps 和 SRE 领域，这样做是有充分理由的。AI technologies have, for the right reasons, started being leveraged in the DevOps and SRE areas of the production support, troubleshooting, maintenance, and upgrade operations.

• 智能体式编程 / 自主软件工程
  Agentic Coding / Autonomous Software Engineering
• 智能体可观测性
  Agent Observability
• 推理优化与模型路由
  Inference Optimization and Model Routing
• 小型语言模型（SLM）
  Small Language Models (SLMs)
• 推理模型
  Reasoning Models
• 可解释人工智能
  Explainable AI
• 虚拟现实（例如 VR/AR/MR/XR）
  Virtual Reality (eg VR/AR/MR/XR)
• 认知服务
  Cognitive Services
• 边缘 AI 推理
  Edge AI Inference
• 模型上下文协议（MCP）
  Model Context Protocol (MCP)
• 多模态人工智能
  Multi-modal AI
• AI DevOps/AI 驱动的 SRE
  AI DevOps/AI-powered SRE

早期大众Early Majority

我们将把以下主题归入早期大众类别，因为这些技术已逐渐成为主流，并被各类组织中的软件开发团队广泛采用。We are moving the following topics to the Early Majority category as these technologies have become more mainstream and are being widely adopted by software development teams in various organizations.

上下文工程Context Engineering

在 AI 智能体时代，上下文工程依然重要，并持续发展，以应用于 AI 驱动的解决方案。Arthur Casals 谈到了这一主题，并着重讨论了上下文工程面临的一项挑战：上下文污染。In the age of AI agents, Context Engineering topic continues to be important and evolving to be used in AI powered solutions. Arthur Casals spoke about this with emphasis on one of the challenges in Context Engineering, the Context Pollution.

Arthur Casals：上下文污染实际上是最大的问题，也是现在最值得强调的问题。因为过去一年左右，我们看到整体上大家都在朝着更大的上下文窗口发展，但如果系统无法区分哪些信息有用、哪些是当前有效的、哪些是替代信息，以及哪些是无关、过时或对抗性材料，那么这并不能真正解决问题。这里我想结合一些自己的研究来谈，但总的来说，更多上下文可能会让智能体变得更慢、更昂贵，也更不可靠。实际要问的是：无论上下文窗口有多大，究竟什么内容会进入上下文？来自什么来源？具有什么权限？用于工作流程中的哪个步骤？这样才能避免出现严重的上下文污染。这涉及检索质量、来源分布、权限规则、整体的上下文模型，以及上下文压缩。我的意思是，“上下文”这个词底下其实堆着一大堆问题。这就是我的看法。

Arthur Casals: Context pollution is actually the biggest problem, the problem worth emphasizing now, because we've seen the past year or so movements toward bigger context windows in general, but they don't really solve the issue if the system cannot distinguish what's useful, what's current, what's alternative information from irrelevant, stale, adversarial material. I'm bringing something from my own research here, but in general, more context can make an agent slower, more expensive, less reliable.

The practical question would be, if you have your context window, regardless of the size of it, what enters the context? From which source, with which permission, for what step of the workflow so you don't have this huge pollution. That talks about retrieval quality, source distribution, professions rules, context model in general, context compression. I mean, there's a huge, huge heap of problems underneath the word context. That's how I see it.

此外，与去年的报告相比，我们认为下列更多主题已准备好进入早期大众阶段。Also, compared to last year’s report, we have more topics as listed below, that we think are ready to move to Early Majority status.

• 生成式人工智能/大型语言模型（LLM）
  Generative AI/Large Language Models (LLMs)
• 检索增强生成（RAG）
  Retrieval Augmented Generation (RAG)
• 知识图谱
  Knowledge Graphs
• MLOps
  MLOps
• 向量数据库
  Vector Databases
• 数据契约
  Data Contracts
• 数据可观测性
  Data Observability
• 图数据分析
  Graph Data Analytics
• 合成数据生成
  Synthetic Data Generation
• 开放表格式（OTF）
  Open Table Formats (OTFs)

晚期大众Late Majority

最后，在我们的采用曲线图中的最后一个类别——后期大众——有几个主题是从早期大众类别转移过来的。这些技术如今已被团队全面采用，并成为其核心架构、设计模式和开发流程的一部分。Finally, in the last category in our adoption graph, the Late Majority, we have a few topics that have moved from the Early Majority category. These technologies are now fully adopted by teams and part of their core architecture and design patterns and development processes.

• 自动化机器学习
  Automated Machine Learning
• 图像识别（计算机视觉）
  Image Recognition (Computer Vision)
• 数字助理
  Digital Assistants
• 数据网格
  Data Mesh
• 流处理与分析
  Stream Processing and Analytics
• 推荐引擎
  Recommendation Engines
• 湖仓一体平台主管
  Lakehouses
• 火花
  Spark

从收养关系图中移除的主题Topics Removed from Adoption Graph

我们还想指出，在本期《AI ML 趋势报告》中，我们移除了一些已使用多年的成熟技术。这些技术已成为主流技术，并得到广泛采用，因此无需再在面向读者的采用情况图表中进行跟踪。We also want to note that in this edition of the AI ML Trends Report, we have removed some mature technologies that have been in use for several years now. These technologies have become very much part of the mainstream technologies and adoption, and don’t need to be tracked on our adoption graph for our readers.

其中一些主题涉及成熟技术，如自然语言处理（NLP）和深度学习、NoSQL 数据库、内存数据网格以及大数据（Hadoop）。Some of these topics include mature technologies like NLP and Deep Learning, NoSQL Databases, In-Memory Data Grids, and Big Data (Hadoop).

结论Conclusion

人工智能是一场重大的技术变革。正如我们现在都知道的，近四年来，它并非一时风潮，也不是技术雷达上的短暂闪现，而是会长期存在。它是过去二十多年里最重要的技术浪潮之一。我们应该抓住这场技术变革带来的机遇，而不是畏惧它。想办法利用人工智能工具，将日常生活中的各种工作流程自动化，无论是办公室项目，还是个人生活中的事务。AI is a significant technology shift. As we all know by now, after almost four years, it is not a fad or some temporary blip on the technology radar. It is here to stay. It is one of the major technology waves in the last twenty-plus years. We should take advantage of this technology transformation, not be afraid of it. Find ways to use AI tools to automate any workflows in your day-to-day life, whether it is in the office projects or in your personal life.

团队对明年的一些预测包括：Some of the team’s predictions for the next year include:

• 我们预计美国的机器人公司将涌现更多创新。
  We expect to see more innovations in US-based robotic companies.
• 我们还会看到很多更便宜、速度更快的模型；TypeSafe 风格的模型会非常有意思。
  We will also see a lot of cheaper and faster models; the TypeSafe style of model would be very interesting.
• 我们将开始看到 AI 在能源和基础设施方面受到限制，具体来说，是数据中心选址所需的电力接入、供电容量和冷却条件。到了 2027 年，一些最重大的 AI 事件实际上可能发生在模型之外；重点将更多转向能源、芯片、基础设施和地缘政治。
  We're going to start seeing limitations on AI, in the areas of energy and AI infrastructure in terms of access for electricity, capacity, and cooling data center locations. For 2027, some of the biggest AI stories may actually happen outside the models themselves; it's going to be more about energy, the chips, the infrastructure, and the geopolitics.
• 每个 AI 智能体都会融入某种形式的 Jev 模型。
  Every AI agent will incorporate some form of the Jev model.
• 随着我们逐步摸索如何构建、运营并将所有这些智能体投入生产，通信将越来越多地发生在智能体之间。
  Increasingly, communications are going to be between agents, moving forward as we figure out how to build, operate and bring all these agents to production.
• 我们会看到，许多混合模型会针对定义明确的任务，使用更小、更专业的智能体。最大的区别将体现在其外围系统上。
  We'll see a lot of things using smaller specialized agents for mixed models, for well-defined tasks. The big difference will be in the surrounding system.
• AI将在我们的日常生活中占据更大的比重。我们认为，AI真正有价值、同时仍需要人类智慧的细分领域将逐渐明朗。
  AI will be even a bigger part of our daily lives. We think the niche areas will be settled where AI will be really valuable and where human intelligence is still required.