Scaling Governed AI Beyond PoCs: A Public Sector Agent Architecture with Lakebase and AgentBricks

Many AI pilots succeed, but productionisation remains difficult in regulated environments where governance, auditability, and state management are required. This session is grounded in Skills SA’s Organisational Self-Assessment (OSA) process, a compliance-driven workflow combining structured operational metadata and unstructured free-text submissions. Around 200 assessments are reviewed annually, requiring significant assessor effort (~1,800 hours).We present a multi-agent architecture built on Databricks Lakebase and AgentBricks. A single Databricks App invokes specialised agents via AgentBricks. OSA data is stored in Delta tables, with Lakebase providing operational state management and governance. Agent prompts and scoring rules are versioned as governed data, and AI/BI Genie is used for natural-language queries over operational metadata.Attendees will learn practical patterns for running production AI systems in regulated environments.