I recently attended Build with Gemini Tokyo, a two-day event hosted by Google. https://cloudonair.withgoogle.com/events/build-with-gemini26q3 The event was a hands-on workshop focused on building AI agents using tools such as Antigravity, ADK (Agent Development Kit), and Agents CLI. Rather than simply learning about these tools through lectures, we went through the full development workflow ourselves—from creating an agent locally to deploying it on Google Cloud. One of the main reasons I decided to attend was that I wanted to change the way I approach AI-assisted development. Over the past six months to a year, I’ve increasingly relied on AI to build web applications and conduct data analysis. The tools I use most often are Antigravity and Codex. At the same time, however, my development process had become somewhat repetitive. I would typically start by defining the requirements, break them down into a detailed implementation plan, and then have AI write the code based on that plan. This is fairly close to what is often referred to as spec-driven development. More specifically, I would create an implementation plan in Antigravity, have a more capable model review and refine it, and then pass the finalized plan to another model for implementation. I had become very comfortable with this workflow, and before I knew it, I had been developing in essentially the same way for the better part of a year. There is nothing inherently wrong with this approach—it works well for building software. But recently, I had started to feel that my process was becoming a little too predictable, and I wanted to experiment with a different way of working with AI. That was when I came across Build with Gemini, an event focused on building applications by combining Google Cloud with AI agents. In this article, I’ll share what we actually did at the event, my impressions of working with ADK and Agents CLI, and what the experience made me realize about my own development style. Tools I Learned in This Workshop In this workshop, we mainly used three tools: Antigravity, ADK, and Agents CLI. Antigravity Antigravity is a development environment designed for building software with AI assistance. I normally use Antigravity IDE, but during this workshop, we worked in an Antigravity 2.0 environment prepared specifically for the hands-on sessions. ADK (Agent Development Kit) ADK is a framework for building AI agents. It allows you to define how an agent behaves and which tools it can use, while supporting the broader development workflow, including building, testing, and running the agent. Agents CLI My understanding is that Agents CLI provides Skills designed to support agent development. These Skills make it easier to delegate not only ADK-based implementation, but also tasks such as testing and deploying agents to Google Cloud. Benefits of Using These Tools Together What stood out to me after using them was how much of the development process could be delegated to AI by combining Antigravity, ADK, and Agents CLI. You could start with a simple natural-language instruction such as: “I want to build an agent like this.” From there, AI could handle a large part of the workflow, from implementation and testing to deployment on Google Cloud. This felt quite different from the way I had been developing before, where I would first define detailed specifications and then ask AI to implement each part step by step. With these tools, it felt possible to delegate work to AI in much larger chunks. That was one of my biggest takeaways from the workshop. Realizing the Difference in Development Styles Through Pair Programming After the workshops on Day 1 and Day 2, we had a final session where participants paired up and built an agent together. I paired up with the person sitting next to me, and we decided to build an agent that visualized token usage. In enterprise projects, teams often use cloud environments or enterprise accounts where AI models are billed based on usage. In those situations, there may be a need to track and visualize how many tokens are being consumed. The Person Next to Me Deployed in 10–20 Minutes What surprised me most once we started was how quickly my partner moved. Within about 10 to 20 minutes, they had already reached the point of deploying the application to Google Cloud. I, on the other hand, struggled quite a bit and only managed to deploy mine near the end of the session, about an hour and a half later. We were building fairly similar applications, so why was there such a big difference? When I asked about their approach afterward, the answer was surprisingly simple. They gave Agents CLI a rough instruction such as: “I want to build an app like this.” Then they focused first on getting something working. They deployed it as quickly as possible, and only afterward did they gradually refine the application to bring it closer to what they had originally imagined. I Spent Too Much Time Finalizing the Specifications Up Front I, on the other hand, had brought my usual development style directly into the workshop. I first worked with Codex to define the specifications in detail, created requirements and planning files, and then tried to implement everything according to that plan. Looking back, I think this significantly reduced the benefits of using Agents CLI. Agents CLI already comes with Skills that support implementation, testing, deployment to Google Cloud, and other parts of the development process. Even so, I was using another model to recreate much of that implementation process from scratch. In other words, even though a workflow had already been designed for this environment, I ended up building a separate workflow of my own. I Was Using Agents CLI Without Really Taking Advantage of It After the workshop, I also spoke with one of the engineers who had been supporting us. They explained that there is nothing wrong with taking a spec-driven approach or using Codex to review the plan. However, if you do that, it is better to also provide Codex with the relevant Agents CLI Skills and the latest information about the environment. Otherwise, another model may create a plan based on outdated assumptions, and that plan may not align well with the mechanisms and workflows already built into Agents CLI. In my case, I was technically using Agents CLI, but I was not developing in a way that fully took advantage of what it was designed to do. Next Time, I Want to Deploy First If I work in a similar environment again, I want to change my approach. First, I would set up ADK and Agents CLI. Then I would give them a simple instruction such as: “I want to build something like this.” From there, I would focus on getting a working version deployed to Google Cloud as quickly as possible. Once that was done, I would add features, improve the UI, and gradually bring the application closer to the final version I had in mind. What I realized through this workshop is that development does not always have to begin with defining everything in detail. The approach of getting something working first and improving it afterward can also be extremely effective. Benefits of ADK and Agents CLI I continued experimenting with ADK and Agents CLI the day after the event, and it reinforced just how useful they can be. They Can Handle Large Chunks of Implementation What surprised me most was how much of the implementation they could handle in a single run. Once I gave them an instruction, they could continue working for quite a while, and by the time I checked again, a substantial amount of code had already been written. The results were also more reliable than I had expected. Of course, some corrections were still necessary, but my impression was that they could move a project forward fairly autonomously without requiring detailed instructions for every individual step. They Can Also Handle Google Cloud Tasks Another major advantage is the range of tasks they can handle. Natural-language instructions can be used not only for application development, but also for Google Cloud integration and deployment. For example, they can help deploy applications to agent runtimes or Cloud Run, as well as set up services such as Firestore. Previously, whenever I needed to configure something in Google Cloud, I would open the console, find the relevant service, and work through the settings myself. Since I started using ADK and Agents CLI, however, I have had far fewer reasons to interact directly with the Google Cloud console once the initial environment is in place. Google Cloud Feels Much Less Intimidating Now I had always found Google Cloud somewhat intimidating because of the sheer number of services and the complexity of the interface. But by combining Antigravity, ADK, and Agents CLI, I can now move development forward without having to understand every detail of Google Cloud upfront. After experimenting with this setup, I felt much more confident about using Google Cloud more actively in future projects. For me, the biggest takeaway was not just that ADK and Agents CLI are useful development tools, but that they significantly lowered the barrier to using Google Cloud. About the Workshop Experience and Environment The Build with Gemini event itself was also very well organized. If I get the chance, I would definitely like to participate again. A Workshop Designed So No One Gets Left Behind The workshop began with a lecture, followed by a hands-on session where participants worked along while watching the instructor’s screen. What stood out to me was how carefully paced the workshop was. Rather than rushing ahead, the instructors regularly checked whether everyone had reached the same point before moving on. The workshop felt designed to minimize the chances of participants falling behind midway through. There were also support staff in the room, so whenever I got stuck, I could ask questions right away. That made it much easier to work through the exercises with confidence. The Venue Was Also Very Comfortable The venue was a modern facility in Takanawa Gateway, and snacks and drinks were provided as well. Even though it was a long workshop, the environment was very comfortable throughout the day. We even received some giveaways, so I was genuinely grateful to be able to experience such a substantial program for free. Overall, I was very satisfied not only with the technical content, but also with the organization and support throughout the event. If there is a similar opportunity in the future, I would definitely like to join again. What I Want to Do Next Now that I have had the chance to learn ADK, Antigravity, and Agents CLI, I want to start using them to build AI agents in real projects. Building a Mental Health AI Agent In fact, I have already started developing a web application designed to support mental well-being. The idea is to create an AI agent that listens to users, helps them organize their thoughts in a way similar to counseling or coaching, and supports them in deciding what to do next. It is still under development, but I already have a basic prototype in place. Next: The Zenn × Google Cloud Hackathon My next goal is to participate in a more serious hackathon organized by Zenn and Google Cloud. I want to continue improving the AI agent I am currently building and submit it there. At Build with Gemini, there were some areas where I was not able to take full advantage of Agents CLI. Based on that experience, next time I plan to try a different approach: build something that works and deploy it as quickly as possible first, then improve it afterward. Moving Beyond Local Development to Something Others Can Use Until now, I have often run AI agents locally through the CLI. But if you want other people to actually use an agent as a product, it obviously needs to run in the cloud. There have been several times in the past when I thought, “It works locally, but how do I actually turn this into a web application that other people can access?” By going through the full deployment process to Google Cloud during this workshop, I feel that the barrier to doing that has become much lower. Going forward, I want to make use not only of ADK and Agents CLI, but also Google Cloud services such as agent runtimes and Cloud Run, and challenge myself to build AI products that are not just for my own use, but that other people can actually use as well. About the Workshop Experience and Environment The Build with Gemini event itself was also very well organized. If I have the opportunity, I would definitely like to participate again. A Workshop Designed So No One Gets Left Behind The workshop began with a lecture, followed by a hands-on session where participants worked along while watching the instructor’s screen. What stood out to me was how carefully the sessions were paced. Rather than moving ahead too quickly, the instructors regularly checked that everyone had reached the same point before continuing. The workshop felt thoughtfully designed to make it difficult for participants to fall behind. There were also support staff throughout the room, so whenever I got stuck, I could ask questions right away. That made it much easier to work through the exercises with confidence. The Venue Was Also Very Comfortable The event was held at a modern venue in Takanawa Gateway, with snacks and drinks provided throughout the day. Even though the workshop lasted for several hours, the environment remained comfortable and easy to focus in. We also received some giveaways, which made me even more appreciative of being able to take part in such a substantial program at no cost. Overall, I was very satisfied not only with the technical content, but also with the organization and support throughout the event. If a similar opportunity comes up in the future, I would definitely like to participate again. What I Want to Do Next Now that I have had the chance to learn ADK, Antigravity, and Agents CLI, I want to start applying them to real AI agent projects. Building a Mental Health AI Agent In fact, I have already started developing a web application designed to support mental well-being. The idea is to create an AI agent that listens to users, helps them organize their thoughts in a way similar to counseling or coaching, and supports them in deciding what to do next. It is still under development, but I already have a basic prototype in place. Next: The Zenn × Google Cloud Hackathon My next goal is to participate in a larger hackathon organized by Zenn and Google Cloud. I plan to continue improving the AI agent I am currently building and submit it to the hackathon. At Build with Gemini, there were some areas where I was not able to take full advantage of Agents CLI. Based on that experience, I want to approach the next project differently: get something working and deployed as quickly as possible, then improve it from there. Moving Beyond Local Development to Something Others Can Use Until now, I have often run AI agents locally through the CLI. But if you want other people to use an agent as a real product, it needs to be accessible beyond your local environment. In the past, I often found myself thinking, “It works locally, but how do I turn this into a web application that other people can actually use?” By going through the full deployment process to Google Cloud during this workshop, that step now feels much less intimidating. Going forward, I want to make use not only of ADK and Agents CLI, but also of Google Cloud services such as agent runtimes and Cloud Run. Ultimately, I want to challenge myself to build AI products that are not just useful to me, but that other people can actually access and use as well.