@claude-flow/swarm

V3 Unified Swarm Coordination Module implementing ADR-003: Single Coordination Engine with Hive-Mind Intelligence

Architecture (ADR-003)

This module provides a complete multi-agent coordination system with hive-mind capabilities:

Key Components

@claude-flow/swarm
├── UnifiedSwarmCoordinator ⭐ CANONICAL ENGINE
│   ├── Configurable agent count (default 15, max 100+)
│   ├── Domain-based task routing
│   ├── Parallel execution across domains
│   ├── Multiple consensus algorithms
│   ├── 4 topology types (mesh, hierarchical, centralized, hybrid)
│   └── Performance: <100ms coordination
│
├── QueenCoordinator 👑 HIVE-MIND INTELLIGENCE
│   ├── Strategic task analysis & decomposition
│   ├── Agent capability-based delegation
│   ├── Swarm health monitoring & bottleneck detection
│   ├── 5 consensus types (majority, supermajority, unanimous, weighted, queen-override)
│   └── Learning from outcomes
│
├── AttentionCoordinator 🧠 ATTENTION MECHANISMS
│   ├── Flash Attention (2.49x-7.47x speedup)
│   ├── Multi-Head Attention (8 heads)
│   ├── Linear Attention (O(n) complexity)
│   ├── Hyperbolic Attention (Poincaré hierarchies)
│   ├── Mixture of Experts (MoE) routing
│   └── GraphRoPE (topology-aware positioning)
│
├── FederationHub 🌐 CROSS-SWARM COORDINATION
│   ├── Ephemeral agent spawning with TTL
│   ├── Cross-swarm messaging
│   ├── Federation-wide consensus voting
│   └── Auto-cleanup & heartbeat tracking
│
├── ConsensusEngines 🗳️ DISTRIBUTED AGREEMENT
│   ├── Raft (leader election, log replication)
│   ├── Byzantine (fault-tolerant, 2/3 supermajority)
│   └── Gossip (epidemic protocol for large swarms)
│
└── SwarmHub (deprecated) - Thin facade for backward compatibility

⚠️ Agent Count: Configurable, Not Limited

The default 15-agent architecture is a recommendation, not a hard limit:

// Default: 15-agent V3 architecture
const coordinator = createUnifiedSwarmCoordinator({
  topology: { type: 'hierarchical', maxAgents: 15 },
});

// Scale up: 50 agents
const largeCoordinator = createUnifiedSwarmCoordinator({
  topology: { type: 'mesh', maxAgents: 50 },
});

// Maximum: 100+ agents (DEFAULT_MAX_AGENTS = 100)
const enterpriseCoordinator = createUnifiedSwarmCoordinator({
  topology: { type: 'hybrid', maxAgents: 100 },
});

Quick Start (Recommended)

import { createUnifiedSwarmCoordinator } from '@claude-flow/swarm';

// Create coordinator
const coordinator = createUnifiedSwarmCoordinator({
  topology: { type: 'hierarchical', maxAgents: 15 },
  consensus: { algorithm: 'raft', threshold: 0.66 },
});

// Initialize
await coordinator.initialize();

// Spawn 15-agent hierarchy
const agents = await coordinator.spawnFullHierarchy();
console.log(`Spawned ${agents.size} agents across 5 domains`);

// Submit tasks to specific domains
const securityTaskId = await coordinator.submitTask({
  type: 'review',
  name: 'CVE Security Audit',
  priority: 'critical',
  maxRetries: 3,
});

await coordinator.assignTaskToDomain(securityTaskId, 'security');

// Parallel execution across domains
const results = await coordinator.executeParallel([
  { task: { type: 'coding', name: 'Core Implementation' }, domain: 'core' },
  { task: { type: 'testing', name: 'Security Tests' }, domain: 'security' },
  { task: { type: 'documentation', name: 'API Docs' }, domain: 'integration' },
]);

console.log(`Completed ${results.filter(r => r.success).length} tasks in parallel`);

// Get swarm status
const status = coordinator.getStatus();
console.log('Domain Status:', status.domains);
console.log('Metrics:', status.metrics);

// Shutdown
await coordinator.shutdown();

15-Agent Domain Architecture

The coordinator manages 5 domains with specific agent assignments:

Domain-Based Task Routing

// Route tasks to optimal domains
await coordinator.assignTaskToDomain(securityTask, 'security');
await coordinator.assignTaskToDomain(coreTask, 'core');
await coordinator.assignTaskToDomain(integrationTask, 'integration');

// Get agents by domain
const securityAgents = coordinator.getAgentsByDomain('security');
console.log(`Security domain has ${securityAgents.length} agents`);

// Get domain status
const status = coordinator.getStatus();
status.domains.forEach(domain => {
  console.log(`${domain.name}: ${domain.availableAgents}/${domain.agentCount} available`);
});

Parallel Execution

Execute tasks across multiple domains simultaneously for maximum throughput:

const tasks = [
  { task: { type: 'coding', name: 'Implement Auth' }, domain: 'core' },
  { task: { type: 'testing', name: 'Security Tests' }, domain: 'security' },
  { task: { type: 'review', name: 'Code Review' }, domain: 'support' },
];

const results = await coordinator.executeParallel(tasks);

// Check results
results.forEach(result => {
  if (result.success) {
    console.log(`✅ ${result.domain}: ${result.durationMs}ms`);
  } else {
    console.error(`❌ ${result.domain}: ${result.error?.message}`);
  }
});

Topology Support

Choose the coordination pattern that fits your needs:

Hierarchical (Default)

const coordinator = createUnifiedSwarmCoordinator({
  topology: { type: 'hierarchical', maxAgents: 15 },
});

• Queen agent coordinates domain leads
• Domain leads manage worker agents
• Best for domain-organized V3 architecture

Mesh

const coordinator = createUnifiedSwarmCoordinator({
  topology: { type: 'mesh', maxAgents: 20 },
});

• Limited peer-to-peer connections (max ~10 per node)
• No central coordinator
• Best for distributed workloads under 20 agents

Centralized

const coordinator = createUnifiedSwarmCoordinator({
  topology: { type: 'centralized', maxAgents: 50 },
});

• Single coordinator hub manages all agents
• Simplest to reason about
• Best for small to medium swarms

Hybrid (Best for Scale)

const coordinator = createUnifiedSwarmCoordinator({
  topology: { type: 'hybrid', maxAgents: 100 },
});

• Mesh workers + hierarchical coordinators
• Combines benefits of both patterns
• Best for large-scale enterprise deployments

Consensus Algorithms

Choose how agents reach agreement:

Raft (Default)

const coordinator = createUnifiedSwarmCoordinator({
  consensus: { algorithm: 'raft', threshold: 0.66 },
});

• Leader-based consensus
• Strong consistency guarantees
• Target: <100ms consensus time

Byzantine Fault Tolerance

const coordinator = createUnifiedSwarmCoordinator({
  consensus: { algorithm: 'byzantine', threshold: 0.66 },
});

• Handles malicious agents
• Byzantine fault tolerance
• Higher overhead but more secure

Gossip Protocol

const coordinator = createUnifiedSwarmCoordinator({
  consensus: { algorithm: 'gossip', threshold: 0.66 },
});

• Eventual consistency
• Low overhead
• Best for large swarms

Performance Targets

The coordinator is optimized for V3 performance requirements:

Performance Monitoring

const report = coordinator.getPerformanceReport();

console.log('Coordination Latency:', {
  p50: report.coordinationLatencyP50,
  p99: report.coordinationLatencyP99,
});

console.log('Throughput:', {
  messagesPerSec: report.messagesPerSecond,
  tasksPerSec: report.taskThroughput,
});

console.log('Utilization:', {
  agentUtilization: report.agentUtilization,
  consensusSuccessRate: report.consensusSuccessRate,
});

Backward Compatibility (SwarmHub)

For existing code using SwarmHub, the compatibility layer is maintained:

import { createSwarmHub } from '@claude-flow/swarm';

// ⚠️ DEPRECATED: Use createUnifiedSwarmCoordinator() instead
const hub = createSwarmHub();
await hub.initialize();

// SwarmHub delegates all operations to UnifiedSwarmCoordinator
const coordinator = hub.getCoordinator();

// Use coordinator for advanced features
await coordinator.executeParallel(tasks);

Migration from SwarmHub

// OLD (deprecated)
import { createSwarmHub } from '@claude-flow/swarm';
const hub = createSwarmHub();
await hub.initialize();
await hub.spawnAllAgents();

// NEW (recommended)
import { createUnifiedSwarmCoordinator } from '@claude-flow/swarm';
const coordinator = createUnifiedSwarmCoordinator();
await coordinator.initialize();
await coordinator.spawnFullHierarchy();

Hive-Mind Intelligence (Queen Coordinator)

The Queen Coordinator provides intelligent task orchestration:

import { createQueenCoordinator } from '@claude-flow/swarm';

const queen = createQueenCoordinator({
  swarmCoordinator: coordinator,
  // Optional: connect to neural learning system
  // neuralSystem: myNeuralSystem,
  // memoryService: myMemoryService,
});

// Analyze a complex task
const analysis = await queen.analyzeTask({
  id: 'task-1',
  type: 'security-audit',
  description: 'Comprehensive CVE audit of authentication system',
});

console.log('Task Analysis:', {
  complexity: analysis.complexity,        // 'low' | 'medium' | 'high' | 'critical'
  estimatedDuration: analysis.estimatedDuration,
  requiredCapabilities: analysis.requiredCapabilities,
  suggestedSubtasks: analysis.subtasks,
});

// Delegate to optimal agents
const plan = await queen.delegateToAgents(task, analysis);
console.log('Delegation Plan:', {
  primaryAgent: plan.primaryAgent,
  backupAgents: plan.backupAgents,
  parallelAssignments: plan.parallelAssignments,
});

// Monitor swarm health
const health = await queen.monitorSwarmHealth();
console.log('Health Report:', {
  overallStatus: health.status,
  bottlenecks: health.bottlenecks,
  alerts: health.alerts,
  recommendations: health.recommendations,
});

// Coordinate consensus with 5 types
const decision = await queen.coordinateConsensus({
  type: 'deployment',
  value: { version: '3.0.0', environment: 'production' },
  consensusType: 'supermajority', // 'majority' | 'supermajority' | 'unanimous' | 'weighted' | 'queen-override'
});

Attention Mechanisms

Six attention mechanisms for intelligent agent coordination:

import { createAttentionCoordinator } from '@claude-flow/swarm';

const attention = createAttentionCoordinator({
  topology: coordinator.getTopology(),
});

// Flash Attention - 2.49x-7.47x speedup for long sequences
const flashResult = await attention.coordinate(agents, task, { type: 'flash' });

// Multi-Head Attention - 8 parallel attention heads
const multiHeadResult = await attention.coordinate(agents, task, { type: 'multi-head' });

// Linear Attention - O(n) complexity for very long sequences
const linearResult = await attention.coordinate(agents, task, { type: 'linear' });

// Hyperbolic Attention - Poincaré distance for hierarchies
const hyperbolicResult = await attention.coordinate(agents, task, { type: 'hyperbolic' });

// Mixture of Experts - Route to top-k best agents
const moeResult = await attention.coordinate(agents, task, {
  type: 'moe',
  topK: 3,
  loadBalancing: true,
});

// GraphRoPE - Topology-aware position encoding
const graphResult = await attention.coordinate(agents, task, { type: 'graph-rope' });

Federation Hub (Cross-Swarm Coordination)

Coordinate multiple swarms with ephemeral agents:

import { createFederationHub, getDefaultFederationHub } from '@claude-flow/swarm';

// Get singleton hub or create custom
const hub = getDefaultFederationHub();
// or: const hub = createFederationHub({ maxSwarms: 10 });

// Register swarms
await hub.registerSwarm('swarm-security', {
  coordinator: securityCoordinator,
  capabilities: ['security-audit', 'penetration-testing'],
});

await hub.registerSwarm('swarm-dev', {
  coordinator: devCoordinator,
  capabilities: ['coding', 'testing', 'review'],
});

// Spawn ephemeral agent (auto-cleanup after TTL)
const { agentId } = await hub.spawnEphemeral({
  swarmId: 'swarm-security',
  ttlMs: 300000, // 5 minutes
  task: { type: 'quick-audit', target: 'auth-module' },
});

// Cross-swarm messaging
await hub.sendMessage({
  from: 'swarm-dev',
  to: 'swarm-security',
  type: 'audit-request',
  payload: { module: 'auth', priority: 'high' },
});

// Federation-wide consensus
const vote = await hub.proposeConsensus({
  topic: 'release-v3',
  options: ['approve', 'reject', 'defer'],
  timeout: 30000,
});

// Get federation stats
const stats = hub.getStats();
console.log('Federation:', {
  swarms: stats.swarmCount,
  ephemeralAgents: stats.ephemeralAgentCount,
  messagesSent: stats.messageCount,
});

Advanced Features

Agent Pool Management

// Get domain-specific pool
const corePool = coordinator.getDomainPool('core');
const stats = corePool?.getPoolStats();

console.log('Core Domain Pool:', {
  total: stats?.total,
  available: stats?.available,
  busy: stats?.busy,
});

// Auto-scaling is built-in
// - Scale up at 80% utilization
// - Scale down at 20% utilization

Custom Agent Registration

// Register agent with automatic domain assignment
const { agentId, domain } = await coordinator.registerAgentWithDomain(
  {
    name: 'security-agent-2',
    type: 'specialist',
    status: 'idle',
    capabilities: {
      codeReview: true,
      securityAudit: true,
    },
    // ... other agent properties
  },
  2 // Agent number 2 → security domain
);

console.log(`Registered ${agentId} in ${domain} domain`);

Event Monitoring

coordinator.on('agent.joined', (event) => {
  console.log('Agent joined:', event.data.agentId);
});

coordinator.on('task.completed', (event) => {
  console.log('Task completed:', event.data.taskId);
});

coordinator.on('consensus.achieved', (event) => {
  console.log('Consensus achieved:', event.data.approvalRate);
});

coordinator.on('swarm.initialized', (event) => {
  console.log('Swarm initialized:', event.data.swarmId);
});

API Reference

UnifiedSwarmCoordinator

Lifecycle

• initialize(): Promise<void> - Initialize coordinator
• shutdown(): Promise<void> - Shutdown coordinator
• pause(): Promise<void> - Pause operations
• resume(): Promise<void> - Resume operations

Agent Management

• registerAgent(agent): Promise<string> - Register agent
• registerAgentWithDomain(agent, number): Promise<{agentId, domain}> - Register with domain
• unregisterAgent(id): Promise<void> - Unregister agent
• spawnFullHierarchy(): Promise<Map<number, {agentId, domain}>> - Spawn 15 agents
• getAgent(id): AgentState | undefined - Get agent by ID
• getAllAgents(): AgentState[] - Get all agents
• getAgentsByDomain(domain): AgentState[] - Get agents in domain

Task Management

• submitTask(task): Promise<string> - Submit task
• assignTaskToDomain(taskId, domain): Promise<string | undefined> - Assign to domain
• executeParallel(tasks): Promise<ParallelExecutionResult[]> - Parallel execution
• cancelTask(taskId): Promise<void> - Cancel task
• getTask(id): TaskDefinition | undefined - Get task by ID

Coordination

• proposeConsensus(value): Promise<ConsensusResult> - Propose consensus
• broadcastMessage(payload, priority): Promise<void> - Broadcast message

Monitoring

• getState(): CoordinatorState - Get current state
• getMetrics(): CoordinatorMetrics - Get metrics
• getPerformanceReport(): PerformanceReport - Get performance stats
• getStatus(): {swarmId, status, domains, metrics} - Get comprehensive status

Integration with agentic-flow@alpha

This module can integrate with agentic-flow@alpha for enhanced capabilities:

import { createUnifiedSwarmCoordinator } from '@claude-flow/swarm';
import { AgenticFlowBridge } from '@claude-flow/integration';

// Connect to agentic-flow for enhanced features
const bridge = new AgenticFlowBridge({
  agenticFlow: { version: 'alpha' },
});

const coordinator = createUnifiedSwarmCoordinator({
  topology: { type: 'hierarchical', maxAgents: 15 },
  // Enable agentic-flow features via bridge
  extensions: {
    transport: bridge.getQuicTransport(),     // QUIC 0-RTT
    learning: bridge.getSwarmLearningOptimizer(),
  },
});

Available from agentic-flow@alpha

Roadmap: Future Enhancements

Based on agentic-flow@alpha capabilities that could be integrated:

Priority 1 (High Impact)

• Native QUIC transport with HTTP/2 fallback
• Swarm learning optimizer for auto-topology
• Free P2P provider integration (GunDB, IPFS)

Priority 2 (Medium Impact)

• WASM-accelerated member indexing
• E2B sandbox agent specialization
• Enhanced message types with fuel/memory budgets

Priority 3 (Nice to Have)

• Advanced gossip variants
• CRDT synchronization
• Production-grade Ed25519/X25519 cryptography

Contributing

This module follows ADR-003: Single Coordination Engine. When contributing:

1. All coordination logic goes in UnifiedSwarmCoordinator
2. SwarmHub is a thin facade - no new logic there
3. Domain-based routing should be used for organized hierarchies
4. Performance targets must be maintained (<100ms coordination)
5. New features should integrate via the extensions API

License

MIT

ADR-003 Compliance: This module implements a single canonical coordination engine with hive-mind intelligence, 6 attention mechanisms, federation support, and backward compatibility via facade pattern.