Goal-Directed Analytics Agents on Apache Iceberg

This article explores the design of goal-directed analytics agents that decompose business objectives into dynamic sub-tasks, execute them against live data, and iterate until conclusions are reached. It details the action loop inspired by the OODA loop (observe, orient, decide, act) and explains how Apache Iceberg tables serve as both the data source and a durable state store for agent action logs. The article covers goal decomposition into sub-task types (semantic queries, SQL exploration, comparisons, external lookups, summarization), tool-use patterns, security boundaries for write operations, and a practical example of investigating a revenue drop. It also discusses failure recovery and performance metrics for such agents.