内容来源：https://aibusiness.com/agentic-ai/harnesses-bring-coordination-guardrails-enterprise-ai-agents

内容总结：

随着企业从试验单个AI智能体转向构建涉及多智能体和多模型的工作流，“AI治理框架”（AI harness）这一术语正日益频繁地出现在业界视野中。然而，何为治理框架、它在智能体工作流中处于什么位置、企业何时才真正需要它，这些问题至今仍无定论。

专注于SAP原生AI集成技术的Debcor Engineering公司首席执行官、创始人兼首席架构师加雷斯·德布鲁因在接受采访时，探讨了AI治理框架在企业工作流中正在形成的作用。他指出，治理框架负责协调多个智能体、管理它们与企业系统的交互，并对AI生成的操作提供管控。

德布鲁因解释说，治理框架位于智能体代码与模型之间，负责处理路由、访问控制、上下文管理、评估和审计。这些管控必须在智能体与企业系统交互之前就位。当多个智能体协同工作时，治理框架使协调多个智能体并管理其交互成为可能。

在治理框架如何管理AI智能体与企业系统之间的交互方面，德布鲁因表示，智能体可能访问企业系统以收集决策所需信息，但当它执行操作时——如创建订单、更新客户记录或移动货物——治理框架会施加必要的护栏和管控。智能体也可能需要与多个企业系统交互以完成任务，治理框架负责管理这些交互，帮助确定智能体可以访问什么、可以执行哪些操作以及如何执行。

对于如何向没有接触过治理框架的人解释其作用，德布鲁因用空中交通管制作比喻。他以销售订单进入公司为例：订单可能通过邮件、传真或格式规范的销售订单送达，需要有人充当空中交通管制员的角色，而治理框架正是扮演这一角色。它可以判断“我识别这个订单，可以处理、查找数据并创建订单”，也可以判断“不行，需要走验证检查流程”。验证可能需要多次AI调用，增加成本并需要额外审计和追踪。治理框架确保一切被正确路由并施加适当的检查，最终生成标准化订单进入客户系统。

关于企业何时真正需要治理框架，德布鲁因认为，如果只有一个智能体独立工作，他质疑是否真的需要治理框架。他以自己创建的一个简单OCR应用为例：在会议上拍摄某人胸牌的照片，让智能体立即读取信息并保存到CRM数据库中——这不需要治理框架，因为这是一个轻量且直接的任务。当多个智能体需要协同工作，或其操作需要更多监督和协调时，治理框架的价值才更加明显。

在预制AI治理框架是否正成为企业可行选项的问题上，德布鲁因指出，AI领域正在发生“自建还是购买”的决策。许多企业因为不一定具备自行构建这些系统的专业知识或信心，正在寻找预制AI解决方案。市场上正在出现预制治理框架和其他打包AI选项，但他认为市场尚未成熟。术语也可能具有误导性，因为一个被营销为单一AI智能体的解决方案，实际上可能涉及多个智能体协同工作。他以SAP Sapphire大会上提到的应付账款智能体为例：该工作流实际可能需要10到15个不同的智能体来处理发票处理、信息核查和路由决策等任务。虽然对外呈现为一个智能体，但背后可能有许多智能体和系统在运作，营销与现实之间尚未完全对齐。企业需要了解这些选项的内部构成、不同智能体如何协同工作，以及治理框架在协调它们中扮演什么角色。

德布鲁因从构建和部署AI中得出的最大教训是：AI不是魔法。企业面临着采用AI的巨大压力，正如当年云计算和物联网一样，但领导者需要回归基本面。首先要问：我们想要实现什么？我们想要什么结果，如何衡量？一旦定义了目标和KPI，就可以确定AI系统需要做什么，以及治理框架在其中的位置。治理框架本质上是一组组件的集合，将模型、智能体和其他系统整合在一起以交付结果。不需要一开始就深入技术细节，从业务成果出发，然后分解不同部分，确定治理框架应扮演的角色。归根结底，治理框架通过协调这些不同部分来帮助管理风险和提升效率。

中文翻译：

由谷歌云赞助
选择你的首批生成式AI用例
要开始使用生成式AI，首先应关注那些能够改善人类信息体验的领域。
“harness”的含义仍在演变之中。
2026年9月14日
随着企业从试验单个AI代理转向构建涉及多个代理和模型的工作流，“AI harness”这个术语出现得越来越频繁。供应商也开始将这一概念打包成面向企业客户的产品，但相关术语仍未统一。什么才算是一个harness，它在代理式工作流中处于什么位置，企业究竟在什么时候才真正需要一个harness——这些都仍是悬而未决的问题。
在本问答中，Debcor Engineering的首席执行官、创始人兼首席架构师加雷斯·德布鲁因探讨了AI harness在企业工作流中正在兴起的角色。该公司专注于与SAP的原生AI集成。他解释了harness如何协调多个代理、管理与企业管理系统的交互，并围绕AI生成的操作提供控制机制。
AI harness究竟是什么？它在代理式工作流中处于什么位置？
德布鲁因：Harness位于代理代码和模型本身之间。它负责路由、访问控制、上下文管理、评估和审计。在代理与企业系统交互之前，这些控制措施就必须到位。
当你真正看到代理的价值时，是在你让许多代理协同工作时。Harness正是使协调多个代理并管理其交互成为可能的东西。
Harness如何治理AI代理与企业系统之间的交互？
德布鲁因：Harness协调代理如何执行其工作并与企业系统交互。用“交易”来描述这种交互可能更贴切。
代理可能会访问企业系统，以收集做出决策所需的信息。但当它采取行动时——例如创建订单、更新客户记录或移动货物——harness会施加必要的护栏和控制措施。
代理也可能需要与多个企业系统交互才能完成一项任务。Harness治理这些交互，帮助确定代理可以访问什么、可以采取哪些行动，以及这些行动如何执行。
你会如何向从未接触过harness的人解释它的作用？
德布鲁因：把它想象成空中交通管制。
以一张进入公司的销售订单为例。有人可能通过电子邮件、传真发送它，或者提交一份格式完美的销售订单。基于所有这些情况，我们需要有人充当空中交通管制。而这实际上就是harness所做的。
它可以说：“我识别这个。我可以处理你的订单，查找数据并创建订单。”或者它可以说：“不，我们需要通过验证检查来运行这个。”这种验证可能需要多次AI调用，这会增加成本，并需要额外的审计和跟踪。
Harness确保一切都得到适当路由，并应用正确的检查。其结果是一份标准化订单，随后可以进入客户系统。
企业在什么时候才真正需要一个harness？
德布鲁因：如果你只有一个代理独立工作，我真的会质疑你是否需要一个harness。
例如，想想一个简单的OCR应用，也就是光学字符识别。我做过一个应用，可以在会议上拍下某人胸牌的照片，让一个代理立即读取信息，并保存到我的CRM数据库中。那不需要harness，因为这是一个轻量且直接的任务。
当多个代理需要协同工作，或者当它们的行动需要更多监督和协调时，harness的价值就会变得更加明显。
现成的AI harness是否正在成为企业的可行选择？
德布鲁因：在AI领域，正在发生一场“自建还是购买”的决策。许多企业正在寻找现成的AI解决方案，因为它们并不总是拥有自行构建这些系统的专业知识或信心。
我们正在看到现成harness和其他打包AI方案的发展，但我认为市场尚未成熟。术语也可能具有误导性，因为一个被营销为单一AI代理的解决方案，实际上可能涉及许多代理协同工作。
例如，在SAP的Sapphire大会上，我们听说了应付账款代理。但当你审视那个工作流实际需要做什么时，它可能涉及10到15个不同的代理，处理发票处理、信息核查和路由决策等任务。
所以，虽然它可能被呈现为一个代理，但背后可能有许多代理和系统在工作。这就是我认为营销与现实尚未完全跟上的地方。
我们会看到更多现成解决方案出现，但企业需要理解这些方案内部有什么、不同代理如何协同工作，以及harness在协调它们方面发挥什么作用。
在构建和部署AI的过程中，你学到的最重要的一课是什么？企业领导者在思考AI harness时应牢记这一点。
德布鲁因：AI不是魔法。企业面临着采用AI的巨大压力，就像当年面对云和物联网等技术时一样。但领导者需要回到 basics。
首先要问：我们试图实现什么？我们想要什么结果，以及我们将如何衡量它？
一旦你定义了这些目标和KPI，你就可以确定AI系统需要做什么，以及harness适合放在哪里。Harness本质上是一组组件的集合，将交付该结果所需的模型、代理和其他系统整合在一起。
你不需要立刻深入技术细节。从业务成果开始，然后分解不同部分，并确定harness应发挥什么作用。
从核心来看，harness通过协调这些不同部分来帮助管理风险和效率。
编者注：本采访经过编辑，以提高清晰度和简洁性。

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
The meaning of harness is still evolving.
September 14, 2026
The term "AI harness" has been appearing more frequently as enterprises move from experimenting with individual AI agents to building workflows that involve multiple agents and models. Vendors are also starting to package the concept into products for enterprise customers, but the terminology is still unsettled. What counts as a harness, where it fit in an agentic workflow when does an enterprise actually need one are still all unsettled questions.
In this Q&A, Gareth de Bruyn, CEO, founder and chief architect at Debcor Engineering, which specializes in native Ai integrations with SAP, discusses the emerging role of AI harnesses in enterprise workflows. He explains how they coordinate multiple agents, manage interactions with enterprise systems and provide controls around AI-generated actions.
What exactly is an AI harness and where does it fit in an agentic workflow?
De Bruyn: The harness is what sits between the agentic code and the models themselves. It handles routing, access control, context management, evaluation and audit. Those controls need to be in place before the agent interacts with enterprise systems.
When you truly see the value of agents is when you have many of them working together. A harness is what makes it possible to coordinate multiple agents and manage their interactions.
How does a harness govern the interaction between AI agents and enterprise systems?
De Bruyn: A harness coordinates how an agent performs its job and interacts with enterprise systems. "Transacts" might be a better way to describe that interaction.
An agent might access an enterprise system to gather information it needs to make a decision. But when it takes an action -- such as creating an order, updating a customer record or moving goods -- the harness applies the necessary guardrails and controls.
An agent might also need to interact with multiple enterprise systems to complete a task. The harness governs those interactions, helping determine what the agent can access, what actions it can take and how those actions are carried out.
How would you explain the role of a harness to someone who hasn't worked with one before?
De Bruyn: Think of it as air traffic control.
Take, for example, a sales order coming into a company. Someone might send it by email, by fax or as a perfectly formatted sales order. Based on all of this, we need someone acting as air traffic control. And that is really what the harness does.
It can say, "I recognize this. I can process your order, look up the data and create the order." Or it can say, "No, we need to run this through verification checks." That verification might require several AI calls, which can add cost and require additional auditing and tracking.
The harness makes sure everything is routed appropriately and that the right checks are applied. The result is a standardized order that can then go into the customer system.
At what point does an enterprise actually need a harness?
De Bruyn: If you just have one agent working by itself, I'd really question whether you need a harness.
For example, think of a simple OCR application, or optical character recognition. I created one where I could take a picture of someone's badge at a conference, have an agent read the information immediately, and save it in my CRM database. That doesn't need a harness, because it's a lightweight and straightforward task.
The value of a harness becomes clearer when multiple agents need to work together or when their actions require more oversight and coordination.
Are pre-made AI harnesses becoming a viable option for enterprises?
De Bruyn: There's a build-versus-buy decision happening with AI. Many enterprises are looking for pre-made AI solutions because they don't always have the expertise or confidence to build these systems themselves.
We're seeing the development of pre-made harnesses and other packaged AI options, but I don't think the market is mature yet. The terminology can also be misleading because a solution marketed as a single AI agent might actually involve many agents working together.
For example, at SAP's Sapphire conference, we heard about an accounts payable agent. But when you look at what that workflow actually needs to do, it could involve 10 to 15 different agents handling tasks such as processing invoices, checking information and routing decisions.
So, while it may be presented as one agent, there can be many agents and systems working behind the scenes. That's where I think marketing and reality haven't fully caught up.
We're going to see more pre-made solutions emerge, but enterprises need to understand what is inside those options, how the different agents work together and what role the harness plays in coordinating them.
What is the biggest lesson you've learned from building and deploying AI that enterprise leaders should keep in mind when thinking about AI harnesses?
De Bruyn: AI is not magic. There's a lot of pressure on enterprises to adopt AI, just as there was with technologies such as cloud and IoT. But leaders need to bring it back to the basics.
Start by asking: What are we trying to achieve? What outcome do we want, and how will we measure it?
Once you define those goals and KPIs, you can determine what the AI system needs to do and where a harness fits in. A harness is essentially a collection of components that brings together the models, agents and other systems needed to deliver that outcome.
You don't need to get into the technical details right away. Start with the business outcome, then break down the different pieces and determine what role the harness should play.
At its core, the harness helps manage risk and efficiency by coordinating those different pieces.
Editor's note: This interview has been edited for clarity and conciseness.