Replacing Multi-Agent Pipelines with Knowledge Graph Control Planes

Moving from Fragmented Agents to Centralized Reasoning

The team initially built a multi-agent system that mimicked a human analyst by assigning separate agents to signal detection, source localization, driver attribution, and synthesis. This architecture failed because context was lost during handoffs, and no single agent possessed the end-to-end business domain knowledge required to produce coherent recommendations. The team abandoned this topology by observing how Claude Code operated in an empty directory, which revealed that a single agent could manage complex tasks by dynamically spawning sub-agents only when specific, focused lookups were required.

Implementing Deterministic Workflows and Control Planes

The rebuild separated deterministic tasks from agentic reasoning. Statistical signal detection now runs as a pre-processing pipeline, ensuring the agent investigates validated anomalies rather than guessing or detecting noise. The system uses a pharma knowledge graph as a control plane rather than a simple lookup table. Every edge in the graph represents a testable hypothesis. The agent traverses the graph, evaluates data against these edges, and iterates until it identifies the root cause. This bounded search space prevents the agent from hallucinating relationships and ensures the final output remains grounded in business logic.

Key Architectural Changes

• Offload deterministic workflows: Move statistical signal detection and thresholding to a standard pipeline that runs before the agent initiates.
• Consolidate reasoning: Use a single primary agent to own the end-to-end reasoning process, delegating specific sub-tasks to dynamic sub-agents as needed.
• Treat graphs as control planes: Map domain entities and KPI relationships into a knowledge graph that dictates valid investigation paths and hypothesis testing.
• Iterative validation: Implement a loop where the agent queries the graph for neighborhood relationships, evaluates the hypothesis against raw data, and traverses further based on supporting evidence.