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

Source: [AI Engineer](https://www.youtube.com/watch?v=u6jJcIFDLE4)  
Feed7 permalink: https://feed7.dev/p/why-we-killed-our-multi-agent-pipeline-subbiah-sethuraman-and-abhilash-a-0fmz3z3  
Published: 2026-07-23T05:00:02.000Z  
Trust: Source Linked (source_linked)

## Why Included

A fixed chain of specialist agents lost context and produced incoherent recommendations. The replacement separates deterministic detection, gives one agent end-to-end ownership, and uses subagents only for bounded investigations.

## Source Summary

ZS first split pharma analysis across signal, localization, attribution, and synthesis agents. Each could produce a plausible fact, yet repeated handoffs lost context and left **no agent owning the full conclusion**.

## Practical Implication

Move deterministic work such as anomaly detection outside the LLM, then wake one reasoning agent from a queue. Let that agent launch **focused subagents** for evidence gathering while retaining judgment, and constrain investigations with a domain knowledge graph.

## Agent-Ready Context

ZS first split pharma analysis across signal, localization, attribution, and synthesis agents. Each could produce a plausible fact, yet repeated handoffs lost context and left **no agent owning the full conclusion**.

Move deterministic work such as anomaly detection outside the LLM, then wake one reasoning agent from a queue. Let that agent launch **focused subagents** for evidence gathering while retaining judgment, and constrain investigations with a domain knowledge graph.

The reported architecture produced analyst-like work in **20–30 minutes after 50-plus turns**, versus three or four weeks for the compared human process. That is one domain-specific account, not a general benchmark for single-agent designs.

## Context Map

- Layer: agent
- Domains: data
- Topics: multi-agent, subagents, harness-engineering

## Uncertainty

- The reported architecture produced analyst-like work in **20–30 minutes after 50-plus turns**, versus three or four weeks for the compared human process. That is one domain-specific account, not a general benchmark for single-agent designs.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.