The Shift to Agentic Engineering

Modern AI development for small teams is moving away from complex, manual infrastructure toward 'vibe coding'—using AI-assisted development tools to implement features through natural language prompts. This approach allows developers to focus on business logic while delegating the heavy lifting of schema updates, vector generation, and API integration to AI agents. The core philosophy is to treat the database as the central intelligence hub, bridging the gap between LLMs (like Gemini) and existing application data.

To move beyond simple full-text search, developers can implement semantic search by leveraging vector embeddings. This process involves:

  1. Vectorization: Converting unstructured data (descriptions, images, audio) into vector embeddings.
  2. Database Integration: Storing these vectors in a managed database (e.g., AlloyDB) that supports vector indexing.
  3. Hybrid Search: Combining traditional keyword-based search with semantic vector search. This ensures that the system understands context (e.g., that 'pet friendly' is equivalent to 'safe for dogs and cats') while maintaining the precision of traditional SQL queries.

The Role of MCP Servers and Tools

Model Context Protocol (MCP) servers act as the bridge between AI agents and your internal data or APIs. Instead of hard-coding every interaction, developers can expose specific tools through an MCP server. When a user asks a question, the agent evaluates its available tools, selects the most relevant one (e.g., get_applicable_promotions), and executes it. This modular approach allows teams to control exactly what data the agent can access and how it interacts with the underlying infrastructure.

Build-Time vs. Run-Time AI

  • Build-Time (Developer Assistance): Using AI to write code, update schemas, and optimize performance during the development phase. The end-user experience remains standard, but the development velocity increases significantly.
  • Run-Time (Agentic Experience): Building applications that feature an agentic interface for the end-user. This requires robust frameworks (like LangChain or Agent Development Kit) to handle real-time user queries and tool invocation.

Key Takeaways

  • Start with the Database: Use managed databases that have built-in AI capabilities to avoid the overhead of maintaining separate vector stores and complex ETL pipelines.
  • Leverage Vibe Coding: Use AI-assisted IDEs to generate implementation plans and code for database schema changes and frontend updates.
  • Prioritize Security: When using agents, use MCP servers to strictly control the scope of data and tools the AI can access.
  • Don't Abandon SQL: Combine traditional structured SQL queries with vector search to get the best of both worlds (precision and semantic understanding).
  • Iterate with AB Testing: When introducing AI features, keep the 'classic' functionality alongside the new AI-powered version to validate performance and user behavior.

Notable Quotes

  • "There's pretty much a myth that you do need a big team, that you need specialized software... The really good way to do it is to identify a few places where your data is, and to try to add AI capabilities on top of that."
  • "We have started moving from vibe coding to agentic engineering... you want to make sure that when you're working with agents, that you are building skills that will empower the agent to achieve the very specific tasks."
  • "Whatever vector database you use, it should also be a good database... don't lose track of or lose sight of what any database really needs to give you, whichever database you choose to go with."