AI Agent Bites #5: Agent Development Kit – Customer Service Agent walkthrough

Motivation

We saw this customer service agent demo in the Google Cloud Next 2025 Keynote. Looked interesting and we wanted to try it out!

Google has provided the “adk-sample” repo to showcase their latest Agent Development Kit, containing various AI agent “samples” built using the ADK. We went ahead and tried out the Customer Service sample, to see how it compares to the keynote demo.

You are given a role of the customer in this sample agent, so we decided to test how it can help revive a neglected orchid from my apartment corridor.

This agent was tested using the ADK web UI, which is part of the Agent Development Kit. The UI provides the means to locally; chat with the agent, manage chat sessions and see past “events” which include a visual breakdown of the tool calling made by the agent. The UI also includes artifacts (for uploaded files) and the “eval” tab for testing various scenarios, but neither of those seemed to work in our case. The screenshot above also shows the customer profile object, which I edited to personalize the experience!

Walkthrough

After instructing it to add the suggested products to the cart, we asked about the price and discount. The agent wasn’t able to give us the price of the cart, but gave us a discount!

Conclusion

As you can infer, the agent wasn’t quite what we saw in the keynote demo—something Patrick Marlow himself acknowledged.

Notably, some functionality like the Live API was missing. When we tried using live video or audio (as shown in the keynote), an error appeared indicating the default model didn’t support the Live API. Switching to a model that does support it led to a different error instead.

Finally, the Agent Development Kit offers a solid development environment with a built-in Web UI for testing. However, more work is needed to bring keynote functionality to production. The Web UI is strictly local, and deploying the agent to Google Cloud’s Vertex Agent Engine means local tools won’t carry over—they must be manually configured. Likewise, the Web UI is only for local testing; a custom interface is needed for production.

Looking to leverage modern AI tools within your company? Get in touch and explore next steps.

Share on: