Skip to main content
The haus²⁵ curation system implements LangGraphJS StateGraph patterns adapted for production efficiency and cost optimization. While inspired by LangGraph’s hierarchical multi-agent coordination, the implementation prioritizes practical deployment needs over complex graph compilation.

​
LangGraphJS Foundation

​
StateGraph Inspiration

The system draws from LangGraphJS StateGraph concepts but simplifies implementation for production requirements: Traditional LangGraphJS Pattern: StateGraph import, createSupervisor with agents array and LLM model, workflow compilation with MemorySaver checkpointer and InMemoryStore. haus²⁵ Adapted Pattern: Direct PlannerSupervisor class with shared agents (RAG, Research, Memory, Blockchain) and specialized agents (Title, Description, Pricing, Schedule, Banner), simplified coordination without compilation overhead.

​
Supervisor Orchestration Pattern

​
Hierarchical Coordination

​
Coordination Implementation

Context Preparation Phase: 
// Phase 1: Shared context preparation (LangGraph state management)
private async prepareSharedContext(eventData: any, userAddress: string) {
  const [userContext, researchData] = await Promise.all([
    this.rag.indexUserHistory(userAddress),
    this.research.researchCategory(eventData.category)
  ]);
  
  return { userContext, researchData };
}
Agent Coordination Phase: 
// Phase 2: Parallel agent execution with shared context
private async orchestratePlan(eventData: any, userAddress: string) {
  const sharedContext = await this.prepareSharedContext(eventData, userAddress);
  
  // LangGraph-inspired parallel agent execution
  const [title, description, pricing, schedule, banner] = await Promise.all([
    this.titleAgent.generateTitle(eventData, undefined, sharedContext.researchData),
    this.descriptionAgent.generateDescription(eventData, undefined, sharedContext.researchData),
    this.pricingAgent.generatePricing(eventData, undefined, sharedContext.researchData),
    this.scheduleAgent.generateSchedule(eventData, undefined, sharedContext.researchData),
    this.bannerAgent.generateBanner(eventData, undefined, sharedContext.researchData)
  ]);
  
  return this.structurePlan(title, description, pricing, schedule, banner);
}

​
Simplified Graph Compilation

​
No-Compilation Approach

Unlike LangGraphJS which requires explicit graph compilation, haus²⁵ uses direct method orchestration: LangGraphJS Approach: createSupervisor compilation with checkpointer and store, runtime execution via compiledApp.invoke with messages array. haus²⁵ Simplified Approach: Direct PlannerSupervisor with generatePlan method, orchestratePlan execution, state tracking, and error handling without compilation overhead.

​
Multi-Level Hierarchy

​
Scope-Based Supervision

CurationCoordinator: Top-level coordination with three supervisors (Planner, Promoter, Producer). Scope-based routing where Promoter inherits Planner capabilities plus promotional content, Producer includes full pipeline coordination.

​
Memory Management Pattern

​
LangGraph-Inspired Persistence

Short-Term Memory (equivalent to LangGraph checkpointer): ShortTermMemory class with sessionCache Map, setAgentState and getAgentState methods using session:agent keys with timestamps. Long-Term Memory (equivalent to LangGraph store): LongTermMemory class with MemoryAgent integration, persistPlanState and retrievePlanHistory methods using on-chain iteration system.

​
Performance Optimizations

​
Context Efficiency

Context Optimization: ContextOptimizedSupervisor with contextCache Map, 5-minute cache expiry, userAddress:category cache keys, and fallback to fresh context preparation when cache misses.

​
Cost Reduction Strategies

Token Optimization:
  • Shared context preparation: Single research call shared across all agents
  • Model selection by task: Gemini Flash for creative, Gemini Lite for analytical
  • Result caching: 12-hour cache for category research data
  • Context filtering: Agent-specific data extraction to reduce token usage