This is a submission for the Google AI Agents Writing Challenge: Learning Reflections
I already knew about the possibility of using multiple AI agents to design complex workflows: my company is building one right now — and I’m partially involved in the development process. But I had never tried Google’s Agent Development Kit (ADK) before. Here’s what my 5-days journey with Kaggle was like.
Agents Intensive
I have no doubt that the most interesting part of the course was the white papers. Despite only lasting five days, they will give me the opportunity to delve deeper into all the topics covered in the coming weeks or even months. References may require even more. Will I have time to read them all?
The Kaggle Notebooks, which I already knew, were an excellent laboratory for putting into practice what I had learned: obviously the first thing I did was install the ADK locally. So, having a free Google AI Studio API key, I did some testing from the terminal emulator using Gemini.
On that note, I think I’ll stick with a local copy of Gemma 3 in the future, as I have no intention of subscribing to a paid plan. This is because my use is exclusively educational, as I’m an employee. Not because I don’t think the service is valuable: but I can’t afford it right now.
Day 1 • Introduction to Agents
I used to think of agentic AI in the singular. I don’t think this is just a problem for my company, but just as DevOps is concerned with optimizing cloud computing costs, so we have been asked to do with AI. So, as it has been from 2018 to today, I had never considered the idea of integrating multiple agents into the same flow.
Technology is evolving: just consider the move from the Completion to the Responses API for ChatGPT. The introduction of the prompt object reduces the cost of system prompts, yet many have not adapted to the new infrastructure to date. Google’s solution goes exactly in this direction: optimization and efficiency.
Among other things, during this study day I learned above all how to use agents sequentially or in parallel. Additionally, I figured out how to use one agent to judge another or orchestrate a group of them. I didn’t think that, regardless of the tool used, it could be so useful.
Day 2 • Agent Tools & Interoperability with Model Context Protocol (MCP)
This was a slightly different day than the others, as it wasn’t focused on using multiple agents. In fact, not even the final project requires using more than one, but out of five days, that was the standard practice for at least four. Here we took a step back, returning to talk about tools and MCPs.
Tools are the reason why I reevaluated the adoption of LLM models in the company. In essence, thanks to them we are able to fill the gap that agents have: knowledge of the present and the future. More or less advanced, models are trained on knowledge bases established at the time they are developed.
You can’t ask ChatGPT what the weather will be like tomorrow like you can with Alexa, because it simply can’t know. Tools allow agents to query external, up-to-date and predictive sources to the model. The MCP protocol is a connector that allows agents to query third-party tools for the same purpose.
Day 3 • Context Engineering: Sessions & Memory
Agents have no memory. This is a problem for two reasons: first, economic, and second, functional. Economically, as I mentioned when switching from Completion to Responses APIs, this means that the entire history of a conversation needs to be submitted to the agent each time. A waste of tokens.
The functional problem is the same. Without history, or context, an agent can’t go beyond the single request it receives. Try thinking about pronouns: if you start a sentence with a pronoun, but the person you’re talking to doesn’t know the previous sentence, what you’re saying makes no sense.
Therefore, Google’s ADK offers two different ways to manage agent memory: one based on a single session and one based on an external database. This solution allows you to save tokens and offer agents a context that is not provided by default. In customer care it’s crucial.
Day 4 • Agent Quality
Test-Driven Development (TDD) should guide any development. It doesn’t. It’s well known that automated testing is the least of developers’ worries — I envy anyone who thinks otherwise. It will probably be the same with agents, but the solutions proposed by Google are still more than satisfactory.
It’s not just a matter of testing. Ensuring the quality of responses also means submitting the solution proposed by an agent to a review: and here we go back to day one. The needs are many, but we can reduce them to two. Quality assurance and error debugging.
Nothing different from what we would do with traditional software development. Languages and environments change, but not the paradigm. There are numerous configurations of agents, models, etc., but the way to subject them to quality testing is always the same.
Day 5 • Prototype to Production
The first thing you should think about is CI/CD. It doesn’t matter what the programming language, framework, or development environment. The same goes for agents. So far, I’ve assumed we’re talking about Python, but Google has made the ADK accessible for use with Go and Java as well.
Everything I’ve talked about so far applies to the Agent2Agent (A2A) Protocol. An ecosystem that allows tools and agents to communicate with each other, regardless of the model used and the domain in which they were instantiated. A kind of MCP for agents: instead of retrieving information and tools for users, it does it for them.
This brings us to the concept of AgentOps. Sooner or later, going back to day one, we’ll have agents that update themselves. Until then, deployment operations will require human intervention: this means that specialized process management professionals will be needed.
Capstone Project
This is completely optional, but I decided to participate because, being ADHD, it gives me the opportunity to focus on a project for the next sixteen days. Honestly, I don’t care about finishing in the top three. I just need to be able to create something from start to finish.
I could choose between four different categories: I opted for Enterprise Agents, because at my age I really struggle not to think about business opportunities when I’m developing something. I work for a company that specializes in customer service, and it’s the industry I know best.
Well, to be honest, it would be more intriguing to throw myself into something different than usual, but I have limited time and for once I want to get something done. I have several options and I need to commit to the most useful one. I think I’ve already identified which one it is, but I’ll let you know when I’m done.
Top comments (0)