Why We Killed Our Multi-Agent Pipeline — Subbiah Sethuraman and Abhilash Asokan, ZS Associates

System Architecture & Workflow Optimization - Separation of deterministic data processing from agentic systems, using statistical methods, guard rails, and thresholds for automated signal detection prior to agent execution [10][11][25] - Consolidation of the core reasoning process into a single primary agent that retains end-to-end ownership, while allowing dynamic launching of sub-agents solely for focused investigation tasks [12][13][26] - Adoption of a lighter architecture derived from observing cloud code operations rather than strictly mimicking human analyst behavior [9][13][25] Domain Knowledge & Knowledge Graph Integration - Construction of a comprehensive knowledge graph mapping out pharmaceutical commercial domains, including geographic entities, payers, accounts, brands, and hierarchical KPIs [16][17] - Utilization of the knowledge graph as a control plane rather than a simple lookup layer, where every graph edge represents a testable hypothesis to bound and guide agent investigation [18][19][21][26] Business Performance & Operational Efficiency - Identification of specific commercial scenarios where a brand's total prescriptions (TRX) drop by 18% over a 4-week timeframe due to insurance coverage shifts to lower tiers [2][3][19] - Significant reduction in analysis turnaround time, enabling the completion of complex analytical tasks in 20 to 30 minutes compared to the traditional 3 to 4 weeks [24]