How do I build an AI Agent using the Vertex AI Agent Engine?

All answers to this question.

The Vertex AI Agent Engine uses a "Reasoning Engine" (based on LangChain) to decompose a complex user request into smaller, actionable steps. To connect it to your SQL database and refund API, you use "Tools." You define your API as a tool in the Agent's configuration, and the model (Gemini) decides when it needs to call that tool based on the user's intent. The Google Agent Development Kit (ADK) simplifies this by providing a framework to build, test, and deploy these agents. It effectively bridges the gap between the LLM's natural language and your company’s technical backend systems.

   Answered 2024-09-14 by Barbara Miller

Are you worried about the agent accidentally issuing too many refunds? You might want to implement a "Human-in-the-loop" step where the agent prepares the refund but requires a manager to click "approve" in a UI before the API call is actually executed.

   Answered 2024-09-15 by Thomas Wilson

• Thomas, that is a brilliant safety suggestion. We were definitely worried about "hallucinated refunds!" Barbara, the "Reasoning Engine" sounds exactly like what we need. I’ll start by looking at the ADK documentation to see how to wrap our existing Python functions as tools. If we can combine the reasoning of Gemini with a manual approval step for financial transactions, this will be a massive win for our support team's productivity.

     Commented 2024-09-16 by Michael Brown

Vertex AI Search is also integrated here. You can have the agent search your help docs if it can't find an order number, giving the user a helpful fallback.

   Answered 2024-09-17 by Susan Taylor

• Susan's right. Using Vertex AI Search as a "retrieval tool" makes the agent much more versatile for non-transactional queries.

     Commented 2024-09-18 by Barbara Miller