Why We Killed Our Multi-Agent Pipeline: Lessons From Pharma Commercial Intelligence

Key takeaways: A practical design principle for agentic systems in regulated, high-stakes domains: derive the architecture from agent behavior, don't impose it. Concrete patterns the audience can apply this week — domain knowledge graphs as agent context, deterministic preprocessing as a complement to agentic reasoning, reference-based context management. An honest case study from production: what worked, what didn't, and the open architectural questions we're still working on. Abstract : We lead the architecture and AI engineering org behind ZS Associates' commercial intelligence platform for pharmaceutical brand teams. The product has two surfaces: a proactive alert system that delivers signal-driven intelligence packets when a brand's KPIs move, and a conversational analytics chat where business users ask ad-hoc questions. A year ago we built both surfaces as separate V1 stacks. They broke in different ways. The diagnosis was the same: we had decided on the structure before we knew what the agent actually needed. This talk is about the design principle that came out of rebuilding both — and what it produced. The architecture is derived, not designed. We stopped trying to predict what scaffolding the agent would need and started designing the system around what the agent's behavior, on real production tasks, actually demanded. Tools, context, structure, and guardrails get introduced at the points where the agent's reasoning needs them — and nowhere else. What that produced is an architecture that's smaller than V1, not bigger. A single agent owns each investigation end-to-end across both surfaces, launching parallel sub-agents when the work needs them — not according to a pre-defined topology. A pharmaceutical commercial knowledge graph — HCPs, accounts, payers, territories, brands, KPIs and the relationships between them — gives the agent the domain context it needs without prompt-engineering heroics. Statistical signal detection runs deterministically before the agent wakes up, so the agent's job is to explain signals, not find them. Raw query results stay out of the context window through a reference-pattern that lets the agent reason over data without drowning in it. Each of those decisions came from watching an agent struggle on a real task and asking what does it need here? — not from sketching the architecture in a doc and forcing the agent into it. The patterns generalize. If you're shipping agents over messy enterprise data — finance, supply chain, claims, operations — the failure modes and the fixes will look familiar. We'll close with the open questions and the pieces we haven't solved yet.