December 6, 2025

• Uncategorized

In this post I’m going to describe the process I went through to create a growth mindset coach using Google’s ADK. For more on why I’m doing the project see this post. I also already built the coach once using a low-code platform, n8n.io, and described that here.

In addition to wanting to learn how to use ADK and code agents in Python, there are a number of practical reasons why I wasn’t fully satisfied with my n8n.io version of the Growth Coach:

1. I wanted the flexibility to integrate the coach with things that didn’t already have built in tools on the n8n platform. For example, he will eventually be integrated with myFitnessPal and Garmin to track my diet and fitness; neither of which have been integrated to n8n (it should be noted that that’s consistent with their enterprise focus). With the flexibility of MCP Servers and tools, it’s now possible to integrate anything with an API to your agents if you’re using code. I did not want to be constrained by n8n’s tool library.
2. I wanted the tighter integration with Google Search. I think because Google has efficiently indexed the entire internet, they are likely to be the winners in this AI arms race and I wanted to get comfortable with their platform.
3. While you can create multi-agent flows in n8n, I think it begins to get cumbersome. Given my background as a software developer I am more comfortable with complex projects having separate files for each agent, having code that can be version controlled, etc… This is true even for the little set of agents I intend to produce, but it’s VERY true for the set of agents that my colleagues at Kyndryl are creating. They’re building agentic flows that do things like decompose millions of lines of COBOL code into user stories that can be rebuilt.

Setting up ADK was surprisingly easy. If you want to follow along at home, I had to install the Google Cloud SDK and CLI, create a project that had access to the API, get logged in to the CLI and authenticated to the Cloud Project. I had the whole thing running on my linux desktop in a few hours. On my windows laptop I used the Git Bash to reproduce it and only had to modify the steps slightly.

The source code for the application I built is on my git repo. There’s a lot of information there about how the code works. My primary observation while building the application was how nicely vibe coding works when you’re doing agentic coding. I used it to create a tool that integrates with the Google Sheets API (sheet_utilities.py) for example that would have been very tedious to code myself. This ability to generate things like that quickly meant that I didn’t miss the GUI interface offered by platforms like n8n.

Google also includes a very nice feature for a built in local web interface that you can use to chat with your newly created application (one of the next steps for me is going to be to look at how to publish this and integrate it with the other agents I want to build). It gets you using and iterating on your application quickly. It also allowed me to record this little video to show you what the agent looks like:

Some of my next steps are going to be to integrate additional agents (like my diet and fitness tracking) and to create new agents. I’m also trying to work out how I can communicate with my agents on my phone.