Multi-Agent Swarms

Multiple specialized agents collaborate on complex tasks.

Overview

Multi-agent swarms enable multiple AI agents with different specializations to collaborate on solving complex problems.

Key benefits:

  • Better quality on complex tasks
  • Multiple perspectives on problems
  • Consensus-based decisions
  • Dynamic role assignment

How It Works

  1. Role assignment: System assigns roles (researcher, engineer, critic, synthesizer)
  2. Parallel execution: Each agent works on the problem from their perspective
  3. Contribution collection: All agent outputs are gathered
  4. Synthesis: Final agent combines insights into unified answer

Default roles:

  • Researcher: Gathers facts, constraints, edge cases
  • Engineer: Proposes implementation plans
  • Critic: Identifies flaws and missing requirements
  • Synthesizer: Combines inputs into final answer

Configuration

{
  swarm: {
    enabled: true,
    size: 4,                    // Number of agents
    dynamic_roles: true,        // Auto-assign roles based on task
    max_concurrent: 4,
    agent_timeout_ms: 30000,
    consensus_candidates: 3     // For high-stakes decisions
  }
}

Usage

Enable swarm mode:

curl -X POST http://localhost:8080/v1/chat/completions \
  -d '{
    "model": "phi3",
    "mode": "swarm",
    "messages": [{"role": "user", "content": "Design a REST API for a blog"}]
  }'

Response Structure

{
  "choices": [{
    "message": {
      "content": "Final synthesized answer..."
    }
  }],
  "swarm_metadata": {
    "agents_used": 4,
    "roles": ["researcher", "engineer", "critic", "synthesizer"],
    "contributions": [
      {"role": "researcher", "summary": "Identified key entities..."},
      {"role": "engineer", "summary": "Proposed REST endpoints..."},
      {"role": "critic", "summary": "Noted missing auth..."},
      {"role": "synthesizer", "summary": "Combined insights..."}
    ]
  }
}

Dynamic Role Assignment

When dynamic_roles: true, the system analyzes the task and assigns appropriate roles:

Example: Code review task

  • Code analyzer
  • Security auditor
  • Performance reviewer
  • Documentation checker

Example: Research task

  • Data gatherer
  • Fact checker
  • Summarizer
  • Citation manager

Consensus Mode

For critical decisions, use consensus voting:

{
  swarm: {
    consensus_candidates: 3  // Generate 3 candidate answers
  }
}

Multiple synthesizers vote on the best answer.

Use Cases

  • System design: Multiple architectural perspectives
  • Code review: Security, performance, style, documentation
  • Research: Comprehensive information gathering
  • Decision making: Consensus on important choices

Performance

  • Latency: 3-5x single agent (parallel execution)
  • Quality: 20-40% improvement on complex tasks
  • Token usage: N x base usage (N = number of agents)