Search Multi-Agent Architecture

Overview

This is a hierarchical three-agent architecture designed for information search and question answering tasks. The architecture follows a hierarchical structure with distinct layers: a control layer (Verifier Agent) that orchestrates the execution, and an execution layer (Search Agent and Answer Agent) that performs the actual work.

Hierarchical Architecture

The architecture is organized into two distinct layers:

Layer 1: Control Layer (Verifier Agent)

• Role: Acts as the control layer, evaluating information sufficiency and routing requests to the appropriate execution layer agent
• Layer Position: Top layer - always executes first
• Input: Original question + all historical search queries + all retrieved information + context
• Output: <verify>yes</verify> or <verify>no</verify>
• Routing Logic:
  • yes: Information is sufficient → Route to Answer Agent (Layer 2)
  • no: More information needed → Route to Search Agent (Layer 2)

Layer 2: Execution Layer

The execution layer consists of two specialized agents that perform the actual work:

2a. Search Agent

• Role: Generate search queries to gather external information
• Layer Position: Execution layer - triggered by control layer
• Trigger: Only executes when Verifier Agent (Layer 1) outputs no
• Output: <search>query</search> format
• Behavior: Only generates search queries, does not attempt to answer questions directly

2b. Answer Agent

• Role: Generate high-quality answers when information is sufficient
• Layer Position: Execution layer - triggered by control layer
• Trigger: Only executes when Verifier Agent (Layer 1) outputs yes
• Output: <answer>final answer</answer> format
• Behavior: Synthesizes all available information into a comprehensive answer

Execution Flow

The hierarchical execution follows a top-down approach:

Layer 1 (Control):
  Verifier Agent → Evaluates if historical information is sufficient
    │
    ├─ If "no" → Layer 2 (Execution): Search Agent → Generates search query → Return search query
    └─ If "yes" → Layer 2 (Execution): Answer Agent → Generates answer → Return answer