Monzo 自建编码智能体走红：把内部平台变成智能体的基础设施

简要描述Brief Description

Felipe 和 Francesco 介绍了 Monzo 如何基于现有的工具、知识来源和沙箱化编排平台，构建内部编码智能体，并将其应用于工程和业务工作流。Felipe and Francesco explain how Monzo built an internal coding agent on an existing platform of tools, knowledge sources and sandboxed orchestration, then reused it across engineering and business workflows.

目录Table of Contents

• 简介
  Intro
• 数字解读 Monzo
  Monzo by the numbers
• 走红的 Slack 代理
  The Slack agent that went viral
• 10% 的合并 PR，每天 1,800 项任务
  10% of merged PRs, 1,800 tasks a day
• 大语言模型 = 一大堆小乘数
  LLMs = lots of little multipliers
• MCP 网关
  The MCP gateway
• 沙箱中的智能体编排
  Agent orchestration in a sandbox
• 作为构建模块的智能体
  The agent as a building block
• 事件模式与值班
  Incident mode and on-call
• 非工程师发布更改
  Non-engineers shipping changes
• 为什么还要继续自己开发它
  Why keep building it in-house

引言Intro

我们是 Felipe 和 Francesco。今天我们来谈谈使用 LLM 构建 Monzo 自有智能体所带来的复合价值。We are Felipe and Francesco. We are here to talk about the compound value of building Monzo’s in-house agents using LLMs.

数字看 MonzoMonzo by the numbers

Monzo 是一家数字银行，致力于让每个人都能更好地使用金钱。为了让你了解编码智能体所面临的问题规模，我们的工程师团队维护着 3,000 多项后端服务、约 13,000 个数据仓库模型，以及几十种机器学习预测器。因此，我们做各种各样的事情，比如评估信贷、判定金融犯罪，以及为应用、后端和数据处理非常复杂的用例。所以我相当肯定，即使就在今天，在这场会议之外，你也听说过很多公司采用 AI 智能体进行编码的成功案例。事实上，几周前纽约这里开了一场会议，随后有一篇文章谈到了 2026 年科技行业的现状。文中引用了 X 上的一句话，希望这只是个玩笑：“如果你还没在构建自己的 AI 编码工具链，那你算得上是一家严肃的科技公司吗？”Monzo is a digital bank trying to make money work for everyone. to give you a size of the problem for coding agents, we have our teams of engineers maintaining over 3,000 back-end services, around 13,000 data warehouse models, and dozens and dozens of ML predictors. So, we do all sorts of things like deciding on credit, deciding on financial crime, and doing very complex use cases both for the app, the back-end, and data. So, I'm pretty sure by even by today, and outside of the conference, you have heard a lot of success cases of companies adopting AI agents for coding. in fact, there was a conference a couple of weeks ago here in New York, and this article came out talking about the state the state of the tech industry in 2026. and there is this hopefully jokingly quote from X saying that if you're not building your own AI coding harness, are you even a serious tech company?

所以，值得思考一下，对吧？So, some food for thought, right?

那款走红的 Slack 智能体The Slack agent that went viral

2026年1月，Monzo 内部有一个关于 Agent Chip 的 Slack 频道火了起来。它用了一种小表情符号，这个表情在公司里流行起来是有原因的。你可以直接给它任何提示词，只要对消息点一下这个小表情，它就会把消息当作提示词接收。后来，这种用法不断发展，我们也加大了投入，于是它成了 Slack 和整个工作环境中随时待命的队友。举几个例子，它确实让人感觉像个同事，就在那里。你随时可以 @Agent Chip，让它帮你做点什么、回答问题、了解某些架构，或者只是陪你聊聊天，不过我觉得它不太喜欢聊天。Inside Monzo in January 2026, a Slack channel for Agent Chip went viral. It used a little emoji that became popular inside the company for a good reason. You can give it any prompt directly by just reacting with the little emoji to the message and it will pick it up as a, you know, prompt. so that evolved, we double down, and that became an always available teammate inside our Slack and whole environment. to give you a few examples, it does feel like a coworker. it's just there. You can always at Agent Chip and get it to do something for you, answer a question, find out something about some architecture, or just have a chat with you, but I don’t think it likes that so much.

它也能做代码审查。就像一位资深工程师一样，它会告诉你“我觉得没问题”，然后继续忙自己的事。你也可以给它分配 Linear 任务，也就是给它工单，它就会马上开始处理。 ## 合并 PR 中的 10%，每天处理 1,800 项任务it can also do code reviews. So, like a very senior, engineer, you will tell you it looks good to me and go on your merry way. And you can also assign, linear tasks, as well. So, you can give it tickets and it will just get into it. ## 10% of merged PRs, 1,800 tasks a day

所以，这一切都很好。把时间推进到今天，Agent Chip 编写的代码占整个技术团队所有已合并 PR 的 10% 以上，涵盖数据、机器学习、后端工程、应用等领域。它每天还会运行 1,800 多项任务，这个数量相当可观。而且，业务用户也开始积极采用它了。所以，它不只是用于编程和技术工作，也在尝试回答业务问题、编写查询，或满足业务用户可能有的各种自助需求。 再说一次，你们已经听过很多公司采用 AI 编程代理的成功故事了，但那又怎样？我们今天为什么要来？我们今天想分享什么？难道我们是想吹嘘：Anthropic 发布 Claude Tag 六个月前，我们就已经把它做出来了吗？也不是不可以。但这不是今天演讲的重点。so all of this is good and well. forward into today, Agent Chip authors over 10% of all merged PRs across tech, so that's data, machine learning, that's, back-end engineering, app, etc. Also runs more than 1,800, tasks per day, which is a lot. and it's being actively, starting to be adopted by business users as well. So, it's not just coding and tech, it's also trying to answer business questions, writing queries, or self-serving anything that a business user may need. So, again, you've heard many success stories about companies adopting AI coding agents, but so what? Why are we What are we trying to share today? Are we saying that we want to brag that we built Claude Tag 6 months before Anthropic released it? We could be. That's not the focus of today's talk.

大语言模型 = 一大堆小乘法器LLMs = lots of little multipliers

我们想分享一下 Agent Chip 成功的秘诀，而秘诀就是 LLM。不过实际上，在这里，LLM 指的是“许多小型助推器”（lots of little multipliers）。我们相信，正是 Monzo 强大的平台，让我们能够如此迅速地打造出极具价值的 Agent Chip。所以，当我们说它迅速走红，大家可以写个提示词、用 React 开发，然后盈利时，这背后其实依托的是我们早已具备的一系列能力：告警、监控、与第三方的集成（例如 Francesco 提到的那些）、便于测试的预发布环境、队列抽象、轻松部署、数据库抽象，以及我还可以在这里列举的其他许多能力。归根结底，首先正是我们平台所具备的这些能力，让 Agent Chip 得以成功。we want to share the secret sauce of what makes Agent Chip a successful product and it's LLMs. But actually LLMs in this case stands for lots of little multipliers. We believe that what enabled us to build Agent Chip so quickly and so invaluable way is a very strong platform at Monzo. So when we say that this went viral and you could write a prompt, react, and profit. Actually this comes from a number of capabilities we already had. So alerting, monitoring, integrations with third parties such as the ones Francesco mentioned, a staging environment that allows us to test easily, queuing abstractions, easy deployment, database abstractions, and a number of other things I could be mentioning here. So first and foremost what we have in our platform is what allowed Agent Chip to be successful.

但事情并未止步于此。早在作为远程代理的 Agent Chip 出现之前，我们就已经开始投入资源，支持 AI 编程代理在本地环境中运行。But this does not stop there. We saw even before Agent Chip as a remote agent existed, we started to invest in supporting AI coding agents even in local environments.

MCP 网关The MCP gateway

我们最先做的事情之一，是开发了一个 MCP 网关，让我们能够轻松访问一系列知识源、内部工具，以及 Francesco 提到的一些服务。架构大致如下：我们要连接的对象包括内部服务和知识源。我们将它们置于自建的 MCP 服务器之后，以便确定性地选择要向智能体开放哪些内容，同时管理不同智能体的授权和身份验证。这项工作最初是为了满足本地编码智能体的连接需求。到现在，大家可能已经在多场分享中听到过：我们能够安全地向智能体提供的上下文和数据越多，效果就越好。One of the first things we did was develop an MCP gateway that allowed easy access to a set of knowledge sources and internal tools and some of the services that Francesco mentioned. So this is roughly how the architecture looks like. We have the things we want to connect as the internal services and knowledge sources. And we put that behind the MCP server that we built to get the deterministic behavior choosing what are the things we want to expose to the agents and then we manage also authorization, authentication for the number of different agents. Again, this started in the world of local coding agents that wanted to connect and by now probably everyone has heard in multiple sessions that the more context and the more data we are able to give safely to our agents the better.

沙箱中的智能体编排Agent orchestration in a sandbox

之后，我们构建了代理编排系统，这正是 Agent Chip 在幕后运行的核心，也让它能够在沙箱中运行，并使用受到严格限制的工具和权限。因此，如果我们回到示意图，MCP 服务器的整个架构基本上都浓缩在这个方框里，而我们正是在这里连接 Agent Chip 架构的其余部分。我们让远程代理在隔离命名空间中的沙箱内运行，然后由编排系统触发它们，并从中读取和获取这些信息。这里的核心理念是——而且我们也在多个服务提供商、实验室等机构中看到这种趋势——我们希望将其与生产环境中的其他基础设施隔离开来。再说一次，我们是一家银行，身处受监管的环境中，因此我们要确保代理带来的非确定性受到控制。After that, we built the agent orchestration system, which is exactly what powers Agent Chip behind the scenes and makes it work in a sandbox and with very strict tools and permissions. So, if we go back to our diagram, the whole architecture for the MCP server basically condenses into this box and this is where we connect the rest of the Agent Chip architecture. So, we have remote agents running inside a sandbox in an isolated namespace and then the orchestration system triggers and then reads and gets this information from that. The main idea here is and again, this is coming across in multiple providers and labs and so on is that we want to separate this from the rest of our infrastructure in production. Again, we are a bank, we live in a regulated environment, so we want to make sure that the non-determinism from the agents is controlled.

但与此同时，我们也希望拥有一个环境，让它能够放开手脚运行和发展。再说一次，如果我们把这个展开，就能让你更深入地了解底层正在发生什么。在这个隔离的命名空间中，会启动多个 Pod，并提供智能体运行所需的沙箱。智能体镜像定义了智能体能够使用的具体工具，包括环境、CLI，以及身份验证和授权；后两者同样由网关提供。代理则让我们能够连接到外部世界。例如，Agent Chip 就是通过它连接到我们的代码和代码仓库，并与之协作。当然，我们只会授予它我们选择开放的出站访问权限。这样，我们就能确保访问受到严格控制，并且只会有意地开放我们希望它能够查看的内容。But at the same time, we want to have an environment where we can let it run loose and develop. And again, if we explode this to give you a bit more under the hood visibility of what is going on. So, in this isolated namespace, multiple pods come up with sandboxes that the agents are running on. The agent image is where we define the exact tools the agents will have access. So, things like the environment, the CLIs and the authentication and the authorization that again, we get from the gateway itself. The proxy is what allows us to connect to both the exterior world. So, that's how for instance Agent Chip is able to connect to our code, to our repositories, and work with that. but, of course, we only get the egress that we choose to give it. so, that is how we make sure that it's very controlled and deliberately giving access to the things we want it to be able to see.

最后，那个小小的蓝绿色方框就是智能体本身运行的地方。所以，层层嵌套的那个方框，就是发生非确定性的地方，也是智能体可以执行所有想做的事情的地方。整个图示就是这样：我们可以让 MCP 服务器访问所有与智能体相关的内容，也可以安全地向它开放我们的数据访问权限，而智能体编排系统则负责处理周围的一切。And finally, the little blue-green box there is where the agent itself is running. So, that box inside the box inside the box is where the nondeterminism happens, and it's where the agent can perform everything that it wants. So, this is how the whole diagram looks like when we are able to give our MCP server access to everything that we care about for the agent, we are able to safely give it access to our data, and the agent orchestration system handles everything that's around it.

作为构建模块的智能体The agent as a building block

有了这些构建模块，之后的任何用例都能更轻松地在此基础上构建。正如 Francesco 提到的，我们可以把它连接到 Slack，再连接到 GitHub 和 Linear。我用另一种颜色标出这一点，表示我们还可以把它连接到 3,000 多种服务，而这些服务也能作为构建模块。因此，就像我们平台中的多个组件为构建 Agent Chip 并从中获益提供了支持一样，如今 Agent Chip 本身也成了一个构建模块，我们公司的各个部门也可以利用它来打造其他用例。由此我们开始看到，这些构建模块不仅帮助我们在内部开发了更多产品功能，也带来了切实的业务成果。正如预期的那样，我们观察到工程师能够执行的任务范围更广、数量更多，因为我们不再需要有人时刻守着他们的智能体。With these as building blocks, any use cases that come after it are much, much easier to be built on top of. So, as Francesco mentioned, this we can connect Slack to, and then we can connect GitHub, and then we connect Linear, and then I just use a different color to show this is the more than 3,000 services that we can be connecting to this as a building block. So, in the same fashion that we had multiple things in our platform that enabled us to build Agent Chip and get value out of it. Now that Agent Chip is also a building block, the multiple parts of our bank can also leverage it to create other use cases. So, we start to see this building blocks enabled us to not only get more product features internally, but to drive concrete business outcomes. we observed a higher breadth and volume of tasks that engineers are able to run, as you expect, because then we don't need people to be babysitting their agents all the time.

他们可以使用编码助手处理特定任务，也可以将任务委派给远程代理。我们可以更好地自动化常规流程。这样，人类就能专注于更高价值、尤其是存在模糊性的问题。下面是一些我们能够实现的示例。这是 Agent Chip 正在以事件响应模式工作，并利用 MCP 获取高度相关的数据。They can use their coding assistant for specific tasks, but they can also delegate to remote agents. We can automate routine processes much better. So, that allows humans to focus on the higher value problems where the ambiguity lies particularly. Some examples of these that we are able to do. this is Agent Chip working in incident mode and leveraging the MCP to get very relevant data.

事故模式与值班Incident mode and on-call

每当我们识别并触发一起新事故时，Agent Chip 就会加入频道并开始调查：理解日志和追踪信息，连接到相关的具体服务，当然还会查看代码本身，了解发生了什么，并以高信噪比的方式协助人工处理。我们不希望 Agent 在事故期间说太多，或者不断测试各种假设；我们希望它始终保持高信噪比，同时也让人工决定 Agent Chip 能提供多少支持。我们可以继续在 Slack 上使用它，但不只是通过提及它来触发，还可以在我们所谓的 ask channels 中使用。它本质上是 Monzo 内部小组之间的互动。如果某个小组需要了解后端服务、产品开发或路线图的背景信息，Agent Chip 也可以连接我们的知识来源，为人工参与者起草回复，供其参考并提供支持。Whenever we identify and trigger a new incident, Agent Chip joins the channel and is able to start an investigation, understanding logs, understanding traces, being able to connect to that exact service, and of course, the code itself to see what's going on and to help the human with a very high signal. We don't want the agent to be saying a lot of things during an incident and testing hypotheses like we want it to maintain high signal, and we give the humans also control of how much Agent Chip will be able to support there. We can continue to use it on Slack, but not only by triggering it by mentioning it, but also in what we call ask channels. It's basically squad-to-squad interaction inside Monzo. A squad needs to have some context on a back-end service, on a product development, on the road map, and Agent Chip can also, since it can connect to our knowledge sources, it's able to draft a response for the human in the loop to get as reference and support there.

也许回复本身就足够了，Agent Chip 可以自行解决这个请求工单。有时它会要求提供更多上下文，这时可以标记人在回路中，由他们介入提供更多支持。最近我们一直在做的一件事也创造了很大价值：对数据质量警报进行分诊，并检查可能引发问题的后端变更。基本上，我们把它称为 Agent Chip on Call，它会观察出现的故障，并自动启动调查。这是它的被动响应版本。不过，如果这些错误从一开始就没有发生呢？基于这个想法，我们创建了主动版本，它会监控后端变更，并在下一批次执行处理数据之前，识别这些变更可能对我们的数据造成哪些影响。Maybe the response will be enough, and Agent Chip can just solve that ask ticket by itself. Sometimes it will ask for more context, and the human in the loop can be tagged and come in to support more. something that we're doing very recently that's been creating a lot of value is also triaging data quality alerts and checking for back-end changes that can create problems. So, basically we call it Agent Chip on Call and it observes the failures that we have and triggers an investigation automatically. This is the reactive version of it. But, what if the errors never happened in the first place? We created based on this the proactive version that monitors back-end changes and identifies which are the potential outcomes that can have for our data even before the data is processed in the next batch execution.

以下是我们的技术团队使用它的一些示例。These are examples of how our technology team is using it.

非工程师发布变更Non-engineers shipping changes

除此之外，我们观察到，公司里的非专业人员参与我们系统建设的门槛降低了。通过 Slack 上一个非常便捷的入口，非专业人员可以根据产品反馈、想法或正在进行的讨论，直接调用 Agent Chip，获得一个可用的初稿，而且它已经在预发布环境中通过测试。之后，也只有在这时，他们才会请工程师参与，确保变更得到妥善审核和部署。因此，这并没有降低 Monzo 软件开发的严谨性，但确实让更多非专业人员能够更深入地参与团队工作。Beyond that, we've observed a lower barrier for non-specialists in the company to contribute to our systems. With a very quick entry point for Agent Chip on Slack, non-specialists can, based on product feedback, based on ideas, based on discussions they are having, just call Agent Chip, get a draft that works, that is already tested in staging, and then, and only then, involve an engineer to make sure we get the changes reviewed and deployed properly. So, this doesn't change the rigor to which we develop software at Monzo, but this does allow a number of non-specialists to also start contributing more closely to the team.

为什么要一直坚持内部开发它？Why keep building it in-house

你现在可能会想，市场发展得很快。我们在这场大会上，以及其他场合，都听到了许多令人兴奋的新进展、新产品和新功能。是什么让我们相信，到了 2026 年及以后，Agent Chip 依然有价值？正如我所说的，而且我想再强调一遍，Agent Chip 具有可组合性，它现在是我们平台中的一个新构建模块。因此，它能让 Monzo 很快实现新的智能体应用场景。现在，我们可以在满足所有安全要求以及后端和平台要求的同时，安全地将它们连接起来。我们的平台从定义上来说与模型提供商无关。所以，如果我们回到那张图，那里那个蓝绿色的小智能体方框可以非常轻松地替换成任何模型。You can be thinking now, the market is moving fast. We are hearing a lot of exciting developments and new products and new features coming from multiple companies here in the conference and otherwise. What makes us believe that Agent Chip continues to be a valuable thing at this point in 2026 and moving forward? As I said and I like to reiterate, Agent Chip being composable, it is now a new building block that we have in our platform. So, it very quickly enables new use cases for agentic implementation at Monzo. Now that we can connect it very safely with all our security requirements and all our back end and platform requirements at the same time. Our platform is provider-agnostic by definition. So, if we come back to the diagram, the little blue/green agent box there can be very easily swapped for any model.

我们可以根据具体用途动态选择最合适的模型。因此，从定义上来说，我们并不专属于任何一家供应商，这让我们能够灵活判断哪些模型最符合 Monzo 的需求。最后，我们还在不断拓展所覆盖的开发范围。因此，即使我们今天停止开发，这项工作也已经很有价值；但我们还能覆盖更广泛的场景，比如在后端拟定变更、在应用中拟定变更，在暂存环境中测试这些变更，并覆盖数据测试。谁知道呢？也许甚至还可以加入另一端的预测功能。因此，当远程代理能够端到端地完成这些工作时，我们就能更好地理解并集成我们的系统。市场仍在发展，这并不意味着我们一定会一直依赖 Agent Chip，但它的这些方面以及其他特性，让我们坚信它如今已成为我们内部平台中非常重要的一部分。We can dynamically choose which model best fits the purposes that we have. So, we are not specifically tied to any provider by definition and that gives us flexibility to understand which ones will fit Monzo's needs the best. and finally, we are continuing to move to cover a bigger development window. So, even if we stop development today, this would already be valuable, but we can cover a much broader sense in terms of we draft changes in the back end, we draft changes in the app. We can test these changes in a staging environment, cover the data tests. Who knows? Maybe even predictors on the other side. So, when remote agents are able to do this end-to-end, we can understand and integrate our systems much better. So again, the market is moving, doesn't mean that we are necessarily married to Agent Chip until the end, but these and other aspects of it make us strongly believe that it is a very strong part of our internal platform now.

我希望大家记住的关键要点是：如何从大语言模型中获益——但不是传统意义上的大语言模型，而是你和团队可以利用来放大价值的许多小型倍增器。谢谢。And the key takeaway that I want you all to take away is how can you benefit from LLMs, but not the traditional LLMs, but the lots of little multipliers that you and your team can use to leverage value. Thank you.