docs/adr/001-multi-agent-architecture.md

ADR-001: Multi-Agent Architecture

Status

Accepted

Context

MediGuard AI needs to analyze complex medical data including biomarkers, patient context, and provide clinical insights. A monolithic approach would be difficult to maintain, test, and extend. We need a system that can:

• Handle different types of medical analysis tasks
• Be easily extensible with new analysis capabilities
• Provide clear separation of concerns
• Allow for independent testing and validation of each component

Decision

We will implement a multi-agent architecture using LangGraph for orchestration. Each agent will have a specific responsibility:

1. Biomarker Analyzer - Analyzes individual biomarker values
2. Disease Explainer - Explains disease mechanisms
3. Biomarker Linker - Links biomarkers to diseases
4. Clinical Guidelines - Provides evidence-based recommendations
5. Confidence Assessor - Evaluates confidence in results
6. Response Synthesizer - Combines all outputs into a coherent response

Consequences

Positive

• Modularity: Each agent can be developed, tested, and updated independently
• Extensibility: New agents can be added without modifying existing ones
• Reusability: Agents can be reused in different workflows
• Testability: Each agent can be unit tested in isolation
• Parallel Processing: Some agents can run in parallel for better performance

Negative

• Complexity: More complex than a monolithic approach
• Overhead: Additional orchestration overhead
• Debugging: More difficult to trace issues across multiple agents
• Resource Usage: Multiple agents may consume more memory/CPU

Implementation

class ClinicalInsightGuild:
    def __init__(self):
        self.biomarker_analyzer = biomarker_analyzer_agent
        self.disease_explainer = create_disease_explainer_agent(retrievers["disease_explainer"])
        self.biomarker_linker = create_biomarker_linker_agent(retrievers["biomarker_linker"])
        self.clinical_guidelines = create_clinical_guidelines_agent(retrievers["clinical_guidelines"])
        self.confidence_assessor = confidence_assessor_agent
        self.response_synthesizer = response_synthesizer_agent
        
        self.workflow = self._build_workflow()

The workflow is built using LangGraph's StateGraph, defining the flow of data between agents.

Notes

• Agents communicate through a shared state object (GuildState)
• Each agent receives the full state but only modifies its specific portion
• The workflow ensures proper execution order and handles failures
• Future agents can be added by extending the workflow graph