Agent development and AgentOps with BigQuery, ADK, and MCP

Key Concepts

• Agentic Data Cloud: A vision where AI is integrated across all layers of the Google Cloud stack, from TPUs to BigQuery.
• ADK (Agent Development Kit): An open-source toolkit for building agents with pre-built BigQuery tools.
• MCP (Model Context Protocol): An architectural standard for connecting AI agents to data sources.
• Agent Analytics: A framework for monitoring, logging, and evaluating agent performance using BigQuery.
• Managed MCP Server: A zero-ops, fully managed infrastructure for secure, governed BigQuery connectivity.
• Data Agent: A high-level abstraction where Google handles reasoning and Gemini integration, while users provide business context.

---

1. Building Agents with BigQuery: The Stack

Google presents a tiered stack for building agents, balancing flexibility with built-in value:

• Top of Stack (Data Agents): Offloads reasoning and Gemini integration to Google; users provide context.
• Middle of Stack (Managed/Open-Source): Includes Managed MCP servers and open-source toolkits.
• Bottom of Stack (APIs): Provides maximum flexibility for developers to build custom skills and tools.

2. Development Tooling

Google ADK (Agent Development Kit)

• Functionality: Provides foundational tools (table discovery, metadata retrieval) and advanced tools (forecasting, contribution analysis).
• Efficiency: Developers can build functional data agents in under 10 lines of code.
• Stability: Now available in a stable version for production use.

MCP Toolbox

• Purpose: An open-source solution for connecting agents to data via the Model Context Protocol.
• Key Feature: Supports parameterized SQL, allowing agents to trigger specific, deterministic actions without needing to write raw SQL.

Managed MCP Server

• Infrastructure: A fully managed, scalable, and secure endpoint (bigquery.googleapis.com/mcp).
• Governance: Integrates with IAM, audit logs, and VPC Service Controls.
• Security: Administrators can use IAM deny policies to restrict read-write access, ensuring agents cannot unintentionally modify data.

3. Agent Observability and Analytics

The transition from development to production requires specialized observability:

• Agent Logs: Essential for performance monitoring, error tracking, and "LLM as a judge" evaluations.
• Agent Analytics Plugin: An open-source plugin for ADK and callback handler for LangGraph.
• Real-time Insights: Uses BigQuery streaming writes to log agent activity, allowing for near-real-time analysis of latency, token consumption, and tool usage.
• Multimodal Support: BigQuery natively supports multimodal logs (images, video, audio) via object references.

4. Real-World Application: Carrefour Case Study

Guillaume, a Data Architect at Carrefour, shared their implementation of an internal data agent ("Phoenix Darwin"):

• The Problem: High turnover and repetitive questions regarding the company's massive data platform (150+ pipelines, 600+ consumers) were overwhelming the data team.
• The Solution: An agent integrated into Google Chat that leverages RAG (Retrieval-Augmented Generation) on documentation and chat history.
• Technical Architecture:
  • Hosting: Cloud Run.
  • Inference: Gemini Pro 2.5.
  • Security: Implemented OAuth to ensure agents query data using the end-user's credentials, preventing unauthorized data access.
• Outcome: The agent significantly reduced the burden on the data team. By using BigQuery Agent Analytics, the team identified that documentation was the most queried resource, allowing them to optimize the agent's knowledge base.

5. Key Takeaways and Best Practices

• Strategic Use Cases: Do not apply GenAI to every problem; select use cases where AI provides clear, measurable business value.
• Proximity: Deploy agents where users already work (e.g., Google Chat, Slack) to ensure immediate adoption.
• Integration Realities: Be prepared for the complexity of integrating agents with legacy systems or mainframes that lack modern APIs.
• Continuous Monitoring: Treat agent observability as a core requirement, not an afterthought. Use logs to refine prompts and tool selection iteratively.

"Agent logs are as important as the agent code. Because many times, unless the agent actually went through the execution, you don't know what the agent did." — Sandeep Karmarkar