I built three AI agents in two weeks and one typo nearly broke me

monisha M R — Fri, 21 Aug 2026 12:04:18 +0000

I built three AI agents in two weeks. One typo nearly broke me

I am Monisha, a Computer Science undergraduate in Bangalore and I just finished Google Clouds Gen AI Academy APAC Edition, which was run through Hack2skill. This program had three tracks. I was able to deploy three AI agents. I used a lot of Cloud Shell. I encountered one very stubborn bug. Here is how it all went down.

Track 1: Teaching an AI barista not to lie

My project was to build an AI "barista" for a coffee shop. You ask it for a drink recommendation. It answers based on an actual menu. This sounds simple. The real point of this project was to use something called Retrieval-Augmented Generation or RAG for short. Of putting the whole menu into the prompt every time the AI agent calls a tool that reads a menu file only when it needs to. This keeps things fast and cheap and importantly it keeps the agent honest.

I built this project using Googles Agent Development Kit. I wrapped it in a simple chat interface using Streamlit. Then I deployed it to Cloud Run using a service account that only had the permission it needed to call Gemini. I did not give it any permissions.

The fun part was trying to break it. I asked for something " warm" and the AI agent correctly recommended the Espresso Solo. Then I asked for a matcha frappuccino, which's not on the menu and instead of making something up it politely told me it was not available and suggested some real alternatives. Then I told it I was lactose intolerant and it only recommended dairy- drinks skipping the cappuccino and croissant entirely. All three tests worked like they should. It was really satisfying to watch the AI agent work correctly.

Track 2: Letting an AI agent talk to business data

The second track was about connecting an AI agent to BigQuery through a server so it could query actual structured data on the fly instead of having anything hardcoded. This project was more about the backend work than an interface but it is the part that actually matters if you want an AI agent to do real work with real enterprise data.

Track 3: An AI agent that asks before it touches anything

This project took the longest. It honestly taught me the most. I built a "coffee shop manager" AI agent that reads years sales data, cross-references it against this years graduation ceremony schedule and figures out where staffing is going to fall short.

Two things about this project stuck with me. First the AI agent does not just run code on the actual container. It runs everything inside a Cloud Run sandbox, which is an isolated environment that spins up in milliseconds. So when it writes a Python script to correlate sales spikes with ceremony times that is happening safely boxed off not directly on my infrastructure.

Second the AI agent does not just go ahead. Update my spreadsheet. It analyzes the data tells me where the bottlenecks will be and then asks me if I want it to add these tasks to my list. I said yes and then did it create the tab and write the tasks in. No silent edits, no assuming it knew best. The AI agent waited to be told what to do.

The bug that made me feel dumb for ten minutes

Every real project has one of these bugs. Mine was that my Cloud Run deployment kept failing with a invalid argument" error and something about Artifact Registry. There was no clue anywhere in the logs.

Eventually I just re-read my own deploy command slowly word by word and found the mistake. I had set my region as "us-centrall" of "us-central1". An extra letter "l" of the number "1". One character and the whole thing was quietly pointing at a region that does not exist. I fixed the environment variable, re-ran the same command and it deployed on the first try.

It was a reminder that most of the time the cloud is not broken. I just typed something and did not look closely enough.

What I would actually tell someone starting this

Test your AI agents refusals, not its correct answers. The matcha frappuccino question told me more about whether the RAG setup was working than any of the "correct" answers did.

Do not skip the dedicated service account thing because it feels like extra work. It is five minutes. It means if something ever goes wrong the problem is tiny instead of your whole project.

When something breaks in a deploy command read your own text character by character before blaming the platform. Nine times out of ten it is my mistake.

I have three AI agents up and running all tested and one annoying typo defeated. Now I am moving on to the Ideathon.

Built during Google Clouds Gen AI Academy APAC Edition in partnership, with Hack2skill.

DEV Community: monisha M R

I built three AI agents in two weeks and one typo nearly broke me