Summary
Subbiah Sethuraman and Abhilash Asokan from ZS Associates — a technology firm serving top pharmaceutical companies — deliver a candid post-mortem on why they killed and rebuilt their multi-agent analytics pipeline. Their original system used separate agents for each step of pharma commercial analytics: signal detection, source localization, driver attribution, and synthesis, all connected by an orchestrator agent. While individual agents produced locally correct outputs, the pipeline failed to generate coherent end-to-end recommendations — correctly identifying that patients couldn’t afford a drug but recommending more sales reps rather than addressing payer coverage.
Three root causes emerged: signal detection was non-deterministic (agents sometimes applied statistical methods, sometimes eyeballed the data); context was lost at handoffs between agents; and no agent held full business domain knowledge, including pharma-specific metrics like TRX (total prescription count). The fix came from starting over in Claude Code with a blank directory, only bash and database access, and observing how the system naturally approached the problem.
The key architectural insight: signal detection is a deterministic workflow, not an agentic task. The team separated it into a pure statistical pipeline with guardrails, thresholds, and anomaly detection that feeds a signal queue; a single agent then wakes up to investigate each signal with full context. Parallelism is preserved for data queries, but reasoning is consolidated in one agent rather than distributed across many. The result is a more coherent, debuggable system that avoids the core pitfall of multi-agent architectures: distributing judgment without distributing understanding.
📺 Source: AI Engineer · Published July 23, 2026
🏷️ Format: Workflow Case Study
