Anthropic Teams With Google Cloud to Automate Software Lifecycles

Anthropic 携手 Google Cloud 实现软件生命周期自动化 Claude Code 与 Google Cloud 基础设施的集成，让开发者能够利用并行子代理和 Model Context Protocol 服务器，实现从产品管理到部署的软件全流程角色自动化。 在本次演示中，Ivan Nardini 展示了 Google Cloud 上的 Claude 如何通过采用五种不同的开发者角色，来增强软件开发生命周期的每个阶段。他依次扮演产品经理、UI/UX 设计师、软件工程师、安全工程师和数据分析师的角色，构建、完善、加固并部署了一个实时反馈应用程序。利用 Claude Code 的 Plan Mode、Developer Knowledge API 和并行子代理等功能，Nardini 构建了包含 Cloud Run、Firestore、BigQuery 和 Looker 的云架构。该环节最后进行了现场观众测试，展示了实时数据更新与自动化 AI 情绪分析。 Google Cloud 演示了如何借助 Claude 覆盖软件开发全生命周期的 5 种角色（产品经理、UI/UX 设计、软件工程、安全及数据分析），打通从 AI 代码生成到端到端云端部署的完整闭环。 (0:05) Claude Code 在 Google Cloud 上支持通过应用默认凭据（ADC）实现零密钥快速配置，提供按使用量付费、预留吞吐量、本地化数据安全及高可用性等企业级特性。 (0:39) 在界面与原型设计环节，PM 可直接上传手绘草图交由 Claude Code 生成功能原型，UI/UX 开发者则能利用 Plan Mode 在编写代码前审核并调整逐步实施方案。 (0:51) 借助 Developer Knowledge API 与 MCP 服务器，Claude Code 可直接调用最新 GCP 文档自动设计基于 Cloud Run、Firestore 与 BigQuery 的架构，并启动多个子 Agent 并行完成模块开发与测试。 (11:00) 系统在完成自动化的 OWASP 安全审查与漏洞修补后直接部署至 Cloud Run，现场演示了实时反馈收集、云函数驱动的情感分析，以及结合 BigQuery 与 Looker MCP 的性能监控流程。 (16:50)

Google Cloud Developer Advocate Ivan Nardini demonstrated an end-to-end software development lifecycle using Anthropic's Claude Code on Google Cloud Platform. By adopting five core enterprise personas, Nardini argued that combining specialized agent modes, Model Context Protocol servers, and cloud-native services allows teams to compress prototyping, cloud architecture, security compliance, and telemetry into a unified workflow. Infrastructure Setup and Enterprise Cloud Controls (0:39) Nardini explained that configuring Claude Code on Google Cloud relies on Application Default Credentials (ADC), which detect user credentials and environment variables automatically. An automated setup wizard identifies active projects, determines model availability by region, and pins target models for direct development. This configuration avoids the need to manually rotate or manage API keys across development teams. According to Nardini, deploying Claude within Google Cloud provides specific operational advantages over standalone endpoints. Billing operates on a pay-for-what-you-use basis per token without standard message caps, alongside options to reserve provisioned throughput for enterprise production workloads. Organizational data remains constrained within the customer's Google Cloud project under established enterprise governance policies, while models are served via Google Cloud's regional and global endpoints for high availability. Interface Prototyping and Plan Mode Verification (8:40) Embodying the Product Manager persona, Nardini illustrated how early ideation can bypass protracted design handoffs. Using Claude Code and a foundational `Claude.md` configuration file, an initial application prototype was generated directly from a photograph of a hand-drawn wireframe sketch on a napkin, producing a functional layout within minutes. Transitioning to the UI/UX role, Nardini utilized Claude Code's Plan Mode to refine this initial layout into a three-page interface comprising a landing view, a confirmation page, and a real-time metrics dashboard. Plan Mode enforces a structured checkpoint: Claude produces a step-by-step implementation strategy for all UI components based on simulated Figma design specifications before executing any code. Nardini emphasized that this planning phase allows human developers to inspect, modify, and authorize interface logic before committing changes. Architecture Automation, Sub-Agents, and Security Auditing (11:00) To build the production backend, Nardini integrated Google Cloud Skills and the Developer Knowledge API via a Model Context Protocol (MCP) server. This tool grants the coding agent direct access to current Google Cloud documentation, enabling it to synthesize a serverless architecture without requiring advanced cloud architectural expertise from the user. The resulting application stack deployed a serverless API on Cloud Run, stored raw user responses in Firestore, staged data into BigQuery for analytical processing, and prepared visualizations for Looker. To accelerate delivery, Nardini deployed Claude Code sub-agents in parallel—running simultaneous sprints to generate the API, construct the ingestion pipeline, and configure testing suites concurrently. Prior to deployment, Nardini assumed the Security Engineer role, using a pre-configured Claude Code review workflow to audit the codebase for OWASP vulnerabilities and enforce least-privilege access rules on Google Cloud Service Accounts. The agent identified and patched an authorization vulnerability automatically before pushing the backend API live to Cloud Run. Operational Telemetry and the Agent Registry (20:56) In a live test of the deployed application, audience feedback was captured and displayed instantaneously, recalculating average ratings and response counts. An integrated "feedback analyzer" powered by a Cloud Run function processed incoming qualitative remarks to summarize audience sentiment in real time. Nardini highlighted response submission latency as a primary KPI for application monitoring, demonstrating how operational metrics feed continuous product iterations. To facilitate data analysis without specialized query authoring, Nardini pointed to MCP server integrations with BigQuery and Looker. He directed developers to Google Cloud's agent registry, a centralized directory detailing supported native MCP servers, configuration procedures, observability tooling, and the open-source MCP toolbox on GitHub. These tools allow agents to inspect application databases and configure analytical dashboards directly from standardized prompts. Nardini concluded that combining Claude Code tools—such as sub-agents, skills, and MCP servers—with Google Cloud's managed infrastructure significantly accelerates each phase of software delivery. While time constraints prevented a live BigQuery and Looker query demonstration, Nardini noted that the full source code and setup guides would be made available publicly for independent implementation.

Google Cloud技术推广工程师Ivan Nardini指出，通过将Anthropic的Claude Code代理能力与Google Cloud基础设施深度集成，开发团队能够在贯穿产品策划、界面设计、后端工程、安全审计及数据分析的全生命周期中，依托并行子代理与MCP标准高效交付企业级云端应用。 云端集成配置与企业级底层优势 (0:39) Ivan Nardini介绍了在Google Cloud上配置与运行Claude Code的工程实践。该方案通过应用默认凭据（ADC）实现自动化身份验证，消除了频繁轮换与手动配置API密钥的管理风险；近期上线的配置向导还能自动识别Google Cloud项目及其模型托管区域，便于开发者快速锁定并调用模型。更重要的是，基于该环境的开发保证了业务数据完全停留在用户的Google Cloud安全边界内，契合企业级合规与数据治理规范。 针对生产环境诉求，Nardini对比了标准模型与云端集成的商业与运行差异。Google Cloud提供按Token用量计费的机制，摆脱了单日或单周期对话消息上限的约束；针对高并发或高确定性的生产负载，平台支持预留吞吐量（Provisioned Throughput）。依托Google Cloud覆盖全球与区域的高性能终结点网络，Claude在云上运行兼具高可用性与基础设施级的弹性扩展能力。 原型速绘与计划模式下的界面工程 (0:51) 在产品经理（PM）与UI/UX设计师的协同流程中，Nardini展示了智能代理对传统沟通模式的重塑。针对早期概念沟通成本高的痛点，PM只需提供在咖啡馆餐巾纸上手绘的界面草图照片，并在Claude Code中配合`Claude.md`规范指令，代理便能在数分钟内生成功能性的应用交互原型，显著加快了可行性验证周期。 随后，Nardini演示了UI/UX设计师如何通过Claude Code的“计划模式”（Plan Mode）将粗糙原型转化为生产级交互界面。在该模式下，代理会在落笔编写代码之前先进行架构构思并输出完整实施计划，赋予开发者在编码前审查、调整设计逻辑的控制权。结合通过MCP协议接入的设计文档规范，代理自动生成了包含着陆页、致谢页及实时反馈仪表盘在内的三页面前端应用，确保了视觉组件与业务需求的一致性。 知识API与多子代理协同的架构落地 (11:00) 面对软件工程师在缺乏特定云服务经验时难以部署前端资产的问题，Nardini演示了Google Cloud推出的开发者知识API（Developer Knowledge API）与模型上下文协议（MCP）服务器，配合Google Cloud Skills实现架构落地。该API使Claude Code能实时检索最新的官方技术文档，在缺乏人工干预的情况下自主规划系统架构：选用Cloud Run运行无服务器后端API，利用Firestore承载原始反馈写入，选用BigQuery构建分析数仓，并规划以Looker进行报表可视化。 在代码实现层面，Nardini展示了利用Claude Code的多子代理（sub-agents）并行推进研发。系统同时启动了三个相互独立的子代理，分别负责后端API研发、数据摄取流水线搭建以及前端仪表盘组件开发。类似于敏捷研发冲刺中的多轨推进，各代理并行生成代码并执行自动化测试，大幅压缩了从架构构思到工程就绪的整体耗时。 安全合规扫描与生产环境实时部署 (16:50) 在将应用推向公开环境前，Nardini切换至安全工程师视角执行发布前合规审查。企业级环境下的审查重点在于排查OWASP常见应用漏洞，并对Google Cloud服务账号实施最小特权原则，例如严格限制数据库的读写范围。演示中，Claude Code内置的安全审查工作流自动扫描了代码库，识别出潜在的安全风险并自动完成了补丁修复，随即通过自动化流程将后端API部署至Cloud Run。 部署完成后，Nardini现场展示了运行在Google Cloud上的实时反馈系统。当现场观众提交了满分评分与评语后，系统瞬时重算了响应总量、平均分并刷新图表。演示还触发了运行于Cloud Run上的“反馈分析器”（Feedback Analyzer）函数，该服务能自动抽取用户提交的评论内容，即时提炼并生成针对演讲质量的情感倾向摘要。 基于MCP生态的性能监控与资源拓展 (20:56) 在运维与增长阶段，Nardini以数据分析视角指出，监控系统响应耗时等关键性能指标对优化用户体验至关重要。为打破分析型数仓BigQuery和报表工具Looker的高使用门槛，开发团队可通过MCP服务器让代理以自然语言直接查询底层数据库并构建仪表盘，免去编写复杂SQL或学习专业看板工具的学习成本。虽然受现场时间限制略过了实时交互查询，但核心配置与调用链已完全跑通。 为便于开发者复现，Nardini介绍了Google Cloud推出的代理注册表（Agent Registry）及开源MCP工具箱。注册表集中维护了包括开发者知识服务、BigQuery在内的官方原生MCP服务器，并附带可观测性规范与操作指南；开源仓库则提供了Looker集成的快速配置指引，通过开放生态帮助工程团队将代理能力延展至更深层的数据分析与运维环节。 Ivan Nardini总结表示，Claude Code的核心能力（技能组件、MCP服务器与多子代理）与Google Cloud底层托管服务的结合，已能够贯穿软件开发生命周期的各个角色。但他同时说明，由于现场演讲时间受限，关于BigQuery与Looker的完整数据分析查询流程未能在台上展开，具体工程实现留待开发者参考会后开放的示例代码与官方文档进行探索。

[0:05] Ivan Nardini: wear five different hats [8:40] Ivan Nardini: With the plan mode, what we do, we put Claude in a mode where it thinks... and propose what is going to do before to implement any code. [19:47] Ivan Nardini: Oh, five? Oh, okay. Thank you, man. I really appreciate that. [19:47] Audience member: cool session [22:38] Ivan Nardini: exercise [24:36] Ivan Nardini: just straightforward [24:36] Ivan Nardini: incredible [24:36] Ivan Nardini: Thank you so much, and it was a pleasure being here today.

[0:05] Hello everyone. [0:08] Thank you for being for being here [0:10] today. I'm pretty much excited to have [0:13] this session with you. I know we are [0:15] almost at the end of the event. So thank [0:17] you for bear with us. [0:18] So just to start, let me to introduce [0:20] myself. I'm Ivan Nardini. I'm a [0:23] developer advocate at Google Cloud [0:24] working on building content in [0:26] partnership with Anthropic. And in order [0:29] to start this presentation today, I just [0:31] want to ask a very simple a very simple [0:34] question. So how many of you in the last [0:37] week use any AI tool [0:39] to code or build application? [0:42] Okay, the majority. [0:44] How many of you use the same AI tool to [0:47] build and deploy application on Google [0:49] Cloud? [0:52] Yeah, just a few of them. So the the [0:54] goal here today is just try to make it [0:56] better. So in these in this demo, what [0:59] I'm going to show you is how you can use [1:02] Cloud on Google Cloud to build and [1:05] deploy application [1:06] end to end. And in order to do that, I'm [1:09] going to wear [1:10] five different hats. [1:13] And we will start from [1:15] imagine the use case, right? Imagine the [1:17] scenario. You are an enterprise context. [1:20] In an enterprise context, you probably [1:22] have a team like the one that we are [1:24] visualizing here that [1:27] it's engaged to build a new feature or a [1:30] new product. So starting from the left, [1:33] you probably have a PM. Some of you in [1:35] this room are PMs. And the PM might have [1:38] an idea of how to improve a particular [1:41] product or how to implement a feature. [1:44] And starting from this idea, he share [1:47] this idea with a UI UX design which [1:49] allows him to design the idea, to [1:52] visualize the idea. And after the idea [1:56] start getting a shape, then the idea is [2:00] sent to a software engineer that [2:01] essentially start developing the core [2:04] logic behind this idea in order to ship [2:06] and make the application accessible or [2:09] the feature the future feature [2:11] accessible to everyone. [2:13] And [2:14] but before to ship it, in order to be [2:16] confident on what you are releasing, you [2:18] probably have to pass a security review. [2:21] So that's why you have a [2:23] role like the security engineer that [2:26] will allows you to be confident in the [2:28] release. And finally, once the [2:30] application gets released, [2:33] you probably have a data persona like a [2:37] growth marketer or a data analyst that [2:40] analyze the data that are collected [2:42] through the app in order to generate [2:44] insight and produce feedback to the PM [2:47] in a way that the product or the feature [2:50] itself gets [2:51] improved. So these are kind of you know, [2:55] all the personas that in a simple way we [2:58] can imagine are involved in the software [3:00] development life cycle. Now with respect [3:03] to these personas, Claude could the [3:06] Anthropic's coding agent augment all of [3:09] them providing several components that [3:12] we are going to use [3:13] today. So as I said, in the for the [3:17] remaining part of this presentation, I'm [3:18] going to put the hat of all these [3:20] personas. I'm going to show you how you [3:22] can use these Claude cool components [3:25] together with Claude on Google Cloud in [3:28] order to ship, build and ship a very [3:31] simple a feedback application that at [3:33] the end of the presentation we are going [3:34] to use to rate my my performance here. [3:39] But [3:41] before to start building, of course you [3:43] need to set up and the [3:47] Claude code that we are going to use. [3:49] And I'm so excited and so proud that we [3:52] work together with Anthropic to make [3:54] this process of setting up Code in order [3:57] to use model on Google Cloud in a very [4:00] simple way, in a very straightforward [4:01] way. So, [4:03] you have multiple methods to use cloud [4:06] models on Google Cloud in Code, but the [4:08] simplest one, the faster one is using [4:11] the application default credential, [4:14] which [4:15] automatically [4:16] finds your credential, for example, the [4:18] user one, and based on the environment [4:21] that you're going to use. And as you can [4:23] see in this representation, recently [4:26] Code also introduced this wizard that [4:29] will simply allows you to [4:31] detect your project and your region [4:33] where the models are served, and um [4:37] check which models are available in your [4:39] project, and [4:41] like [4:42] you to pin them in order to start [4:44] building your application. [4:46] At this point, like [4:47] probably you're familiar with this, and [4:49] you're wondering, "Okay, but what's what [4:52] is different to use just Code Code with [4:55] with cloud models? Why using cloud on [4:57] GCP, on Google Cloud?" [4:59] So, there are many reason why you want [5:02] to do why you want to do this. So, first [5:04] of all, because you pay for what you [5:07] use. So, the [5:09] usage of a cloud models on Google Cloud [5:11] is per token, so you don't receive a [5:13] message message cap. And also, if you're [5:17] building a [5:18] enterprise application that needs to go [5:22] to production, you can always access to [5:24] what is called provisioning throughput, [5:26] which essentially will reserve some [5:29] throughput for you in order to build [5:31] this kind of application. [5:33] Um the other important reason why you [5:35] want to consider like cloud on Google [5:38] Cloud is as I said, the setup is pretty [5:41] straightforward using the ADC. You don't [5:44] have API to rotate, or you know, [5:47] environment variable to set in some [5:50] sense. [5:51] So, it's a it's a happy journey in in [5:54] the with respect to this aspect. You can [5:57] access model in your project with your [6:00] own [6:01] you know, policies set and also like the [6:05] data stays in your project while you are [6:07] interacting with the Cloud Code. [6:09] And model are served in multiple [6:12] regions. So, you have global endpoint, [6:14] you have regional endpoint depending on [6:16] you know, the availability that you that [6:18] you need. And as a Google Cloud talking [6:20] about availability, we have very great [6:23] availability service that standards that [6:27] will allows you to [6:29] use Cloud and on one of the most [6:32] performing infrastructure that you can [6:34] find in the in the market. So, these are [6:37] all some of the main reason why you want [6:39] to consider Cloud on Google Cloud [6:41] especially in an enterprise contest. [6:44] So, now that you have like few reasons [6:46] of why using Cloud [6:48] on Google Cloud, [6:49] we are ready to build. [6:51] And so, as I said, I will start wearing [6:53] the hat of a PM. [6:55] So, imagine that you just join the [6:57] company or maybe you're already part of [6:59] the company. You have you you want to [7:02] improve our services, you want to [7:03] implement a new [7:05] features [7:06] with respect to a particular product. [7:08] What it was happening in the past is [7:10] that you have the idea, you go to a UI [7:12] UX designer and you ask him to prototype [7:16] and visualize the idea. Now, with the [7:19] with the Cloud and Cloud Code, all you [7:22] you what you can do is just [7:24] drawing a picture like the one that you [7:26] see here while you're drinking a coffee [7:29] maybe in San Francisco and then let [7:33] Cloud doing [7:34] implementing the idea for you. [7:37] So, let's see this [7:38] in action. [7:43] So, this is a the Cloud Code UI, like [7:47] you will probably familiar with that. [7:48] And you are familiar with the Cloud MD, [7:50] which is essentially give some [7:51] instruction. [7:52] Here we just say um that we are a PM, we [7:56] want us we want to have a starting from [7:58] the picture, we want to render [8:01] a a prototype of the the app, the [8:04] wireframe that we are going to then use [8:07] and pass to the UX designer. And in few [8:10] minutes, you can see how Cloud was [8:12] capable of rendering it and um [8:16] just starting from a very simple uh [8:18] picture or drawing that you you um did [8:22] while you were drinking your coffee. So, [8:25] pretty pretty straightforward. But, [8:27] imagine how much time you save in doing [8:29] this because compared to what you were [8:31] doing in the past with the back and [8:33] forth [8:34] um to in order to get this first [8:36] prototype of your idea. [8:40] Okay. So, at this point, [8:42] the PM gives like uh creates a [8:45] prototypes and pass these prototypes to [8:48] the UI UX [8:50] uh developer. And uh at this point, he [8:52] needs to implement a more solid uh [8:55] interface in order to use it in uh in [8:58] production. So, in in this particular [9:01] use case, what we want to create it's at [9:03] least three like uh pages from the [9:06] landing to the thanking uh [9:09] message like message page and a [9:11] dashboard [9:13] that I would allow me to show you in the [9:15] real time what can be the feedback that [9:17] I will receive from the room. [9:20] So, in this case, [9:21] there are many ways you can you can [9:23] implement this, but in this case, I want [9:26] to use an additional capabilities of [9:28] Cloud Code, which is the plan mode. So, [9:30] with the plan mode, what we do, we put [9:33] Cloud in a mode where it thinks um [9:36] before to um like it thinks and propose [9:39] what is going to do before to implement [9:41] uh any code. And this is very important [9:43] because it gives me like a degree of [9:46] freedom of deciding to change something [9:49] better according to my preference or [9:50] according to some standard that probably [9:53] I will get access through an MCP server [9:55] using uh Figma, for example. [9:59] So, now that we have uh in mind what [10:01] we're going to build, let's see this in [10:03] action as well. So, we started from uh [10:05] the wireframe from the PM. [10:08] Similar prompt. I enable the plan uh the [10:11] plan mode. And so, as you can see [10:13] compared to before, [10:15] um in this case, I'm simulating the [10:17] receiving some instruction from Figma [10:19] using a design doc. But uh as you can [10:22] see compared to before, it creates a [10:24] plan on what it's going to do with [10:26] respect to all the components that are [10:28] defined in the slide. [10:30] We look at them. We are happy. We [10:32] accept. And Copilot will implement all [10:35] of them. And at the end, what we get [10:38] is uh this uh [10:40] optimized version. So, as you can see, [10:42] we start from here and we get this. [10:47] Very very straightforward. But you can [10:49] see how we are shifting from a prototype [10:53] to something that can be used uh in this [10:56] session in a very simple way. [11:00] Okay. So, this is the part that probably [11:02] every view or every of you in this room [11:05] like you do every day, right? [11:08] Um [11:08] let's uh let's wear the third hat, which [11:10] is the one of the software engineer. [11:13] And uh the software engineer [11:17] he receive this uh front end like all [11:20] the components that I was showing you [11:21] before. [11:23] And uh [11:24] maybe it doesn't know anything as [11:26] probably uh some of you in this room. It [11:29] doesn't know how to deploy this [11:30] application on Google Cloud, right? So, [11:36] how you can how you can do that? How you [11:40] can, you know, [11:41] uh hand to have this a clear picture of [11:44] what are the components on Google Cloud [11:46] that you need to use in order to deploy [11:49] a very simple application like the one [11:50] that uh I show you today. [11:53] Luckily, it's not a problem because as a [11:55] Google Cloud, [11:57] we invest a lot of time to integrate [11:59] with these a large uh bycoding ecosystem [12:02] that now is [12:03] um growing around uh [12:06] models. And [12:08] uh we have we in the last few months, we [12:11] released two um important components. [12:14] So, the first one is uh the developer [12:17] knowledge API with this with uh the [12:19] associated MCP server. And the second [12:21] one is the Google Cloud Skills. [12:24] So, with the knowledge with the [12:26] developer knowledge API, you get access [12:28] to the a fresh documentation from Google [12:32] Cloud that can be directly consuming [12:34] cloud code [12:35] uh through the MCP server. And it will [12:38] have cloud code to figure it out what [12:40] what is the best [12:42] uh architecture, what is the best [12:44] implementation uh to deploy a certain [12:46] application on Google Cloud. This is [12:48] very important because again, what what [12:50] we are saying here is that you don't [12:52] need to know uh like how to deploy an [12:54] application on Google Cloud. You can [12:56] just leverage cloud code and this MCP [12:58] server that we expose now on on Google [13:01] Cloud side to build application like [13:03] this one. So, in this case, uh probably [13:07] what we want we are going to do is that [13:08] we are going to deploy the feedback API [13:11] on a serverless function like a Cloud [13:13] Run. We will connect with a web uh web [13:16] um [13:17] oriented DB like a Firestore to corre- [13:20] to collect the raw responses that we we [13:22] will give through the feedback app. And [13:25] then, because we want to have that that [13:27] analytics part, [13:28] we will build uh [13:30] implementation in a way that we can [13:32] store those raw response in an [13:35] analytical data warehouse like BigQuery. [13:38] And we BigQuery we will post process and [13:41] we'll consume this information in a [13:44] dashboard [13:45] uh in a in a in a dashboard like the one [13:48] that you can find in a in Looker. But [13:51] again, it you you can build this [13:54] using [13:56] a Co-Code in combination with the MCP [13:58] server without no without you having a [14:01] prior knowledge of how to do that. [14:04] This, like the MCP server and the [14:07] developer knowledge API that I just [14:08] mentioned, it got paired with also the [14:10] skill part. And the skill part is if [14:14] with the MCP server you will be able to [14:16] design the architecture, with the skills [14:19] you will be able to cover the single [14:20] blocks of this architecture. For [14:22] example, we release a [14:24] a simple skills that [14:26] enabled Cloud to deploy on Cloud Run, to [14:30] deploy an API on Cloud Run, or [14:33] to connect like Cloud Run with [14:36] Firestore. So, it's more about the [14:39] implementation itself rather than, you [14:41] know, giving the overall picture. [14:46] So, once you get Co-Code enabled with [14:50] the MCP server uh of the documentation [14:53] and the skills that I just mentioned, [14:55] you can just directly implement uh the [14:59] architecture that I was showing you [15:00] before and you can do this in parallel. [15:03] So, another components that you can use [15:05] in Co-Code is a sub agents. And as you [15:08] can see here, we are going to spin up [15:10] three different sub agents, one for the [15:12] API, one for the ingestion pipeline, and [15:15] the other one for the dashboard. And you [15:17] can parallelize the implementation of [15:19] each of these components just like you [15:22] run a team a team sprint [15:24] in the in your normal like usual [15:27] development life cycle. [15:30] So, let's see this in action as well. [15:34] So, here we are again. [15:36] First of all, I just want to show you [15:38] that we have enabled the MCP server of [15:40] the documentation and we have the skills [15:43] some of the skills that we pre-built. [15:46] One one time more, I provide a very [15:48] simple prompt. The first step is [15:50] designing the cloud-native back end. So, [15:54] it will start [15:56] like it will provide me [15:58] a draft of the architecture. In this [16:00] case, I could use again the plan mode, [16:02] but for simplicity I didn't. And then he [16:05] used the skill, one of the skill that I [16:07] provide in order to implement the API. [16:09] Let's say that we are happy with the API [16:12] spec, then we have the architecture, we [16:15] have the API spec. The next step is [16:18] running multiple agents in parallel in [16:21] order to implement the [16:24] the three components of the app that I [16:26] was showing you. [16:28] It's it's pretty quick. As you notice, [16:31] GoCode like it also manage the testing [16:34] part after you finish the implementation [16:37] and at the end you will get your [16:39] your app, which is now ready to deploy [16:44] on Google Cloud. [16:49] Okay. [16:50] So, at this point [16:52] we have we have the code of the app that [16:56] is ready to be deployed, but because we [16:58] are deploying on cloud and we want to [17:00] open this up to a larger audience, we [17:02] want to deploy it in a confident [17:04] confidently. Like [17:06] so, [17:08] this is this is when you want to [17:10] consider to run a security review. Now, [17:13] depending on your company, you can you [17:16] can have different security [17:18] requirements. So, for example, you may [17:20] want to check if your [17:22] application is solid with respect to the [17:25] most common [17:27] uh OWASP issue [17:30] or because you are deploying this [17:32] application on cloud and one one thing [17:35] that you need to consider is probably [17:37] you will use what is called a service [17:39] account and you want to limit [17:42] the the service account when you call a [17:45] particular API like the one related to [17:48] reading and writing at DB with respect [17:50] to certain role. So, you are you you're [17:52] sure that [17:54] you're limiting what it can happen when [17:57] when the application runs on operation [17:59] on the on the cloud itself. So, again, [18:02] these are just a couple of examples of [18:04] um [18:05] uh what you want to consider in a [18:07] phase like this one and of course this [18:10] representation is a strong [18:12] simplification of what can happen in the [18:15] in the real life. Um it's just a one of [18:17] the possible scenario that you can have [18:20] and you can see why it's also a [18:21] simplification because we are letting [18:24] the security engineer not only to [18:26] approve if the app is secure enough to [18:29] be deployed, but it will also deploy the [18:30] app in this case. [18:32] But again, this is just a demo so we [18:35] have this degree of freedom. With that [18:37] being said, let's run the final demo and [18:40] let's uh get the app uh running. [18:43] So, [18:44] in this case, as you can see, I use a [18:46] pre-built security review that you can [18:48] find in CoCode. Uh [18:51] very simple um [18:52] very simple prompt also in the in this [18:54] case. [18:55] And what is happening is a CoCode run [18:58] the first task essentially [19:00] double check [19:02] that everything is aligned. It found [19:04] like a possible issue and it [19:07] automatically fixed it. So, in a way [19:09] that the app now is secure and once it [19:12] is secure, you deploy the back uh [19:14] the back end API, and you deploy the API [19:17] itself. [19:18] So, what we get at the end of this, it's [19:21] an end point with our app. The app is [19:23] live, so I will uh [19:27] quickly unlock my laptop, and I will ask [19:33] the backstage to share on my laptop. [19:38] So, the uh the app, as you can see, is [19:40] up and running on GCP. Uh for people [19:43] that doesn't know GCP, this is Cloud [19:45] Run, is a serverless uh [19:47] service that you can use in order to [19:49] deploy app, and uh [19:51] the app at the end looks like this one. [19:54] So, if you remember, I showed you the [19:57] original [19:58] I showed you the feedback uh frame at [20:02] the beginning. So, what we can do now, [20:04] live, I can just give me a score. What [20:07] do you think the session is going so [20:08] far? [20:09] Give me [20:10] >> Oh, five. [20:11] >> Five? Oh, okay. Thank you, man. I really [20:14] appreciate that. [20:16] Okay, cool session. Let's be Let's be [20:19] simple. I saw me, and then [20:22] uh in real time, it updates the number [20:24] of response, the score, and uh you know, [20:27] the visualization. And also, just for [20:29] fun, I built a feedback analyzer. So, [20:32] once I click this, it will run, it will [20:35] call [20:36] Cloud Code [20:37] Cloud on Google Cloud, and uh based on [20:40] the feedback and the comments that you [20:41] sent, it will generate this [20:43] this summary. So, [20:46] pretty pretty straightforward. [20:51] I will ask to go back on the [20:52] presentation. [20:54] Thank you. [20:55] Okay. [20:56] So, at this point, we have the app, we [20:58] collect we collect good feedback. Thank [21:01] you again, man. [21:03] Uh but, the development life cycle is [21:06] still there. Uh we have the last step to [21:08] cover, which is essentially people there [21:11] start using our app. If you saw one of [21:14] the KPI that I had on the dashboard was [21:16] the time the response time. So, how long [21:19] is [21:20] was taking you to just in uh uh like [21:23] providing a feedback. So, these are the [21:25] kind of information that you can use in [21:27] order [21:28] uh to like uh [21:30] through this data uh you can collect [21:32] this data, analyze them, and generate uh [21:35] insights in a way that they can be used [21:37] in order to improve uh the uh [21:39] application. Now, [21:42] running uh if you are new on Google [21:44] Cloud, there are several services that [21:46] you can use in order to analyze these [21:48] data that comes from the app. So, a [21:51] couple of them, one as I said at the [21:53] beginning is an analytical data [21:54] warehouse, which is BigQuery, and uh for [21:57] the reporting part, you have a tool like [22:00] Looker. But again, as we said before, [22:03] you don't need to know how to use [22:05] BigQuery in order to analyze those data, [22:07] as well as how to build a dashboard in [22:10] Looker, because we also provide a MCP [22:13] server for doing this. Now, for the sake [22:16] of time, I'm not going to demo how to [22:18] use an MCP server to query BigQuery or [22:20] building a dashboard, but uh I want to [22:23] quickly share with you [22:25] um where you can find uh this [22:27] information in order to do that after [22:29] this session, because we are going to [22:31] release the code. So, I will ask uh to [22:35] shift back on my laptop. Thank you. [22:38] So, um [22:41] Okay. [22:43] Loading time. So, with respect to all [22:46] the MCP server that are available on [22:48] Google Cloud, we uh recently announced [22:51] the uh agent platform. So, in the agent [22:54] platform, you have uh uh services that [22:57] uh is represented by the agent registry, [22:59] and in the agent registry, you have the [23:01] list of all the MCP server that we [23:04] natively support on Google Cloud. So, [23:06] for example, we have the developer [23:07] knowledge service that I just show you, [23:10] and we also have the BigQuery the [23:14] BigQuery MCP server that you can use [23:18] in order to [23:19] uh [23:20] query the data that we just collected [23:22] from the app. Um it's very like this [23:25] registry is relevant because it tells [23:27] you how to set uh how to set the MCP [23:30] server on your side, as well as it gives [23:33] you some observability feature, and the [23:36] description of all the tools that you [23:38] will find uh for the MCP server. So, you [23:40] know how Clockwork will be able to use [23:43] uh this uh this um this server to uh [23:46] query your data. With respect to the um [23:51] Looker part, also like [23:54] uh we released the MCP toolbox of DB. [23:57] This is a open-source um like a model [24:00] context protocol server, and which [24:02] include an integration with Looker, and [24:05] it's very well um [24:08] we have a very very very well-documented [24:10] quick start on how to set up with uh [24:12] Clockwork and start to using it in order [24:15] to consume that data from BigQuery and [24:17] build your dashboard. So, I leave you [24:19] this as an exercise. Like uh we didn't [24:22] we are going to release the code, so you [24:23] can go home and integrate these two [24:25] parts. It's pretty straightforward, but [24:28] the dashboard that you uh you can [24:29] create, they are pretty they are pretty [24:31] powerful, and you will see uh how nice [24:34] they can be. [24:36] Okay, back to the [24:38] presentation. [24:39] I think I'm just in time. So, time to uh [24:42] time to wrap up. What I tried to explain [24:45] you today [24:46] is essentially uh two things. So, first [24:49] of all, I was trying to and I hope I did [24:51] I did good enough. I was trying to show [24:54] you how like all the components of our [24:56] clock code, including skills, MCP [24:59] server, sub agents, they can really like [25:02] speed up the process of software [25:04] development, as well as how you can use [25:08] clock code with [25:10] cloud models on GCP in a very seamless [25:13] way. Like if you saw we run several [25:15] sessions across like multiple personas, [25:19] and like it was the experience was just [25:23] straightforward. It was just a [25:25] incredible. So, [25:27] this is what you can get if you combine, [25:30] you know, clock code with cloud models [25:33] on on GCP. As I said, the the code [25:38] is going to be available right after the [25:40] session. We have a great quick start, [25:42] and we have a very well maintained [25:45] documentation both on Google Cloud side [25:47] and Anthropic side. So, I highly [25:49] recommend to just go there and check [25:51] out. And then, I hope I covered [25:54] everything, but if you still have [25:56] questions or you want to provide [25:58] additional feedback, just feel free to [26:00] reach out. These are my social media [26:02] accounts. [26:04] point. [26:05] So, with that being said, [26:07] thank you so much, and it was a pleasure [26:09] being here today. [26:21] >> [snorts]