Serverless Expeditions — selected guide
Build a BigQuery AI agent with ADK & Cloud Run
Fourteen episodes on service choice, architecture, scaling, state, security, operations, and delivery
Serverless design and operations through concrete examples: choosing a runtime, ordering events, managing relational state, undoing failed workflows, protecting secrets, and deploying changes safely. The selection also includes two AI-agent examples covering tools and database access.
In this collection Browse 14 summaries 14 of 14
An AI agent can let people ask questions of a database without writing SQL themselves. In this July 2026 demonstration, Martin and his guest build one with Google's Agent Development Kit (ADK) and a managed BigQuery server using Model Context Protocol (MCP), a shared interface for agent tools.
Key points
- The agent queries data rather than searching keywords: Asked for France's top scorers in the 2022 World Cup, the agent uses a goals-scored column and writes SQL to return players from the BigQuery database. The guest argues that reading real tables prevents hallucination; the demo grounds the answer in data, but does not prove the SQL or interpretation cannot be wrong 00:51–02:36(opens in a new tab).
- Instructions explain how to use this database: The system instruction describes tables, columns, query steps, and business rules such as calculating a team's performance rating. The guest compares it to an operating procedure: access to a database does not tell the agent how the business organizes or interprets its data 02:36–04:28(opens in a new tab).
- MCP reduces custom connection code: A hand-written Python tool could also compose and send SQL. The guest favors a managed MCP server because agents can discover its tools at runtime and different agent clients can reuse the same interface, rather than each project building its own adapter 03:23–05:16(opens in a new tab).
- The deployment choices have bounded behavior: The guest presents Cloud Run as a container and service under developer control, and Agent Engine as managing sessions, application versions, and a test UI. He says the Agent Engine setup shown does not expose a REST API directly and needs a client-facing proxy 05:16–06:18(opens in a new tab).
The guest recommends updating instructions as tables and business rules change, and keeping them in separate Markdown files so non-developers can edit them without changing Python code 06:18–08:10(opens in a new tab).
Full video: https://www.youtube.com/watch?v=o3aE4tw3vrI(opens in a new tab)