The Agentic Data Cloud Architecture

Google Cloud is positioning BigQuery as the central hub for an "Agentic Data Cloud," where AI capabilities are integrated directly into the data layer. The strategy focuses on a tiered stack: at the top, managed data agents handle reasoning and context; at the bottom, APIs provide granular control for custom tool building. The goal is to standardize how agents interact with data, eliminating the need for custom, brittle database connectors.

Streamlining Development with ADK and MCP

The Agent Development Kit (ADK) simplifies agent creation by providing first-party tools for BigQuery, AlloyDB, and Google Maps. Developers can build functional agents in under 10 lines of code by leveraging these pre-built tools. For those requiring more architectural flexibility, Google provides an open-source Model Context Protocol (MCP) toolbox, allowing agents to connect to data via parameterized SQL. This approach enhances determinism, as agents pass parameters to predefined queries rather than generating raw SQL, reducing the risk of unintended data modification.

Managed Infrastructure and Governance

To remove the operational burden of hosting MCP servers, Google introduced managed MCP servers. These endpoints integrate directly with Google Cloud’s IAM, audit logging, and VPC Service Controls. Data administrators can enforce security policies—such as denying access to write-enabled SQL tools—at the platform level, ensuring that agent developers cannot bypass governance protocols. This infrastructure is designed to scale, with BigQuery serving as a primary generally available endpoint.

AgentOps: Observability Beyond Traditional Logs

As agents move to production, traditional logging proves insufficient. Agent observability requires capturing multi-modal data, tracking token consumption, and evaluating agent reasoning. Google’s "Agent Analytics" plugin for ADK and LangGraph allows developers to stream agent activity into BigQuery in real-time with a single line of code. Once stored, this data can be analyzed using BigQuery’s AI functions—such as using an LLM as a judge to categorize conversations or evaluate performance—without moving data out of the warehouse.

Case Study: Carrefour's Internal Knowledge Agent

Carrefour implemented an agent to address the "lazy user" problem in their data platform team. With over 600 data consumers and 150 pipelines, the team was overwhelmed by repetitive questions from new hires. By building an agent that ingests documentation and historical Google Chat logs, they automated support. They utilized BigQuery Agent Analytics to monitor user interactions, specifically identifying where conversations stalled, which allowed them to refine the agent’s knowledge base and improve response accuracy over time.