Task Delegation Pattern

The Problem

A single agent can't do everything better than specialized tools. You need a system that selects the right agent for the right task.

Architecture

User Request
     |
     v
[Orchestrator Agent]
     |
     +---> [Research Agent] ----> Web Search, Docs
     |
     +---> [Coder Agent] ----> Code Generation
     |
     +---> [Review Agent] ----> PR Review, Tests
     |
     v
  Final Response

Implementation

1. Intent Classification

The Orchestrator classifies the user request and routes it to the appropriate agent.

2. Routing Matrix

const routes = {
  'code-generation': coderAgent,
  'research': researchAgent,
  'review': reviewAgent,
  'deployment': deployAgent,
  'question': qaAgent,
}

3. Priority Queue

For multiple concurrent tasks: Set priorities (1 = highest). Deadline tracking prevents tasks from waiting forever.

4. Result Aggregation

The Orchestrator collects results from all sub-agents and synthesizes a final response.

Important Aspects

• Timeout: Each sub-agent needs a max timeout
• Retry: Max 2x on errors
• Fallback: What if all agents fail?
• Cost Control: Budget limits per task

Sources

• Agent Architectures (ArXiv)
• Kore.ai Agent Orchestration