Power intelligent agents with AI-native databases

Key Concepts

• Agentic Data Cloud: A vertically integrated data and AI stack designed to act as a "system of action" for autonomous agents.
• AI-Native Databases: Databases (e.g., AlloyDB, Spanner) with built-in AI primitives (vector search, forecasting, graph search) that process multi-modal data natively.
• Model Context Protocol (MCP): An open standard for connecting AI models to enterprise data and tools, facilitating interoperability.
• Text-to-SQL: The process of converting natural language queries into precise SQL commands, enhanced by schema ontologies and query blueprints.
• Parameterized Secure Views: A deterministic security mechanism that restricts agent access to data based on user context, preventing unauthorized data exposure.
• Hybrid Search: A search methodology combining vector search (semantic intent) with full-text search (keyword precision).

---

1. The Agentic Data Cloud Strategy

Google’s strategy shifts from "moving data to AI" (which creates latency) to "moving AI to data." By integrating AI at every layer—compute (TPUs), database (SQL-integrated models), and the agentic layer—Google aims to break down "walled gardens" of siloed data.

• Core Pillars:
  1. AI-Native Databases: Built-in models for structured, numeric, image, video, and graph data.
  2. Data Agent Platform: A framework providing a "nerve center" of enterprise context and natural language interfaces.
  3. Pre-built Agents: Out-of-the-box solutions for builders, operators, and business users.

2. AI-Native Databases: AlloyDB and Spanner

Yannis Papaconstantinou highlighted that modern databases must understand logic, language, and relationships natively.

• AlloyDB: 100% PostgreSQL-compatible, featuring:
  • Vector Search: Uses Google’s proprietary SCANN index (used in YouTube/Search) to support 10+ billion vectors.
  • Performance: Columnar engine technology boosts vector search speed by 4x over standard pgvector.
  • AI Functions: Integration of Gemini-powered functions for forecasting (using TimesFM), sentiment analysis, and summarization.
• Spanner: Leverages global scale for mission-critical AI, specifically utilizing GraphRAG (Graph Retrieval-Augmented Generation) to understand relationships between entities without needing to copy data into a separate graph database.

3. The Data Agent Platform

This platform bridges the gap between natural language and SQL, focusing on accuracy and security.

• Methodology for Accuracy: To reach near 100% accuracy, the platform uses:
  • Schema Ontologies: Describing the database structure.
  • Query Blueprints: Injecting nuanced interpretations for specific questions.
  • Value Searches: Disambiguating questions against actual data.
• Security: Parameterized Secure Views ensure that agents only access data relevant to the specific end-user, preventing prompt injection or unauthorized data scraping.

4. Fireside Chat: The Role of MCP

David Soria (Anthropic) discussed the necessity of the Model Context Protocol (MCP).

• The "N x M" Problem: Without a standard, developers must build custom integrations for every model-to-data connection. MCP provides a universal "connective layer."
• Future Roadmap: The focus is shifting toward serving "skills" (domain-specific knowledge) over MCP servers and enabling "long-running tasks" to support more complex agentic behaviors.
• Quote: "Agents are really good at selecting vast amounts of data... giving them the tools and indexes at their fingertips is a really good way to work through company data." — David Soria.

5. Real-World Applications

• Target.com: Migrated search traffic to AlloyDB, utilizing the SCANN index and dynamic SQL filters. They reported a 20% increase in search relevance and superior performance compared to manual integrations.
• EchoStar: Utilized the Data Agent Kit and MCP to overhaul customer support operations, enabling agents to query maintenance records and analyze machine defects.

6. Pre-built Agents and Developer Tools

Google introduced several tools to accelerate adoption:

• Data Agent Kit: An ecosystem-led developer experience including MCP servers and agent skills.
• Database Observability Agent: Uses Gemini to transform millions of telemetry points into actionable insights.
• Database Testing Agent: Automatically compares performance impacts of configuration or application changes.
• Conversational Analytics Agent: Allows business users to query operational databases using natural language, currently holding the #1 spot on the BIRD benchmark for Text-to-SQL.

Synthesis

The transition to the "Agentic Era" requires databases to evolve from passive storage to active, AI-native systems. By combining AlloyDB/Spanner for high-performance data processing, MCP for standardized connectivity, and Data Agent Platforms for secure, natural language interaction, enterprises can move beyond simple demos to reliable, production-ready autonomous systems. The key takeaway is that the most effective agents are those "grounded" in the enterprise's own semantic knowledge, protected by deterministic security guardrails.