Production-Ready AI Agents: 5 Lessons from Refactoring a Monolith

A technical teardown of refactoring a brittle monolithic AI sales agent ('Titanium') into a production-ready system using Google's Agent Development Kit (ADK). Five key lessons are covered: replacing a monolithic script with a SequentialAgent pipeline of specialized sub-agents, using Pydantic schema injection instead of prompt-based JSON formatting, building a dynamic RAG pipeline with Playwright crawling and Google Cloud Vector Search for hybrid search, adding OpenTelemetry-based observability with live telemetry dashboards, and leveraging ADK's built-in circuit breakers and retry logic to prevent runaway token costs.