Hub-and-Spoke Agent Coordination Research

A Message from the Creator

Hey everyone! Building on Cursor Memory Bank (2,400+ stars), I've been pioneering what's now called Context Engineering with Claude Code.

🔥 The Problem I Discovered:
Agents lose context, forget instructions, and can't coordinate reliably.

🧠 My Context Engineering Solution:
• JIT Context Loading - Agents get exactly what they need, when they need it
• HANDOFF_TOKEN Validation - Confirms agents actually understand
• Hub-and-Spoke Coordination - Eliminates context drift between agents

⚡ What Makes This Different:
Instead of overwhelming agents with everything (which breaks them), my system delivers just-in-time context - precise, relevant information loaded dynamically per task.

🚀 The Result:
A true agent collective that functions as unified intelligence, not just another agent collection.

Come be part of this context engineering breakthrough! 🎯

---

Research: Solving Multi-Agent Coordination

After months of researching and testing agent coordination problems in my Cursor Memory Bank project, I identified three critical failures that plague ALL multi-agent systems. This breakthrough was only possible with Claude Code's sub-agent architecture and hooks system - each agent having its own isolated context window, combined with deterministic hooks for validation, allowed me to reach this next level of coordination research:

🔥 Context Degradation - Agents lose context across interactions and forget critical instructions
🔥 Coordination Drift - Peer-to-peer communication becomes unreliable and non-deterministic
🔥 Quality Inconsistency - Agents skip steps and behave inconsistently without enforcement

Research Hypotheses Under Test

📊 JIT Context Loading Hypothesis

• Testing: Can graph-based context delivery prevent agent focus loss?
• Hypothesis: Dynamic context assembly from knowledge graphs will outperform global context approaches

🔗 Hub-and-Spoke Coordination Hypothesis

• Testing: Can central hub routing eliminate context drift?
• Hypothesis: Routing-agent orchestration will prove more reliable than peer-to-peer communication

🔒 HANDOFF_TOKEN Validation Hypothesis

• Status: PAUSED - Redesigning handoff mechanism
• Issue: Current token validation approach needs refinement for better agent coordination
• Next: Exploring simplified handoff patterns without explicit token requirements

---

Usage

Research framework for reliable multi-agent coordination.

Usage: @routing-agent your request

Core Innovation: Solves fundamental agent coordination problems through hub-and-spoke pattern, automated routing, and mandatory quality gates.

Note: HANDOFF_TOKEN validation is currently disabled while redesigning the handoff mechanism for better reliability.

Quick Start

1. Copy .claude/ directory to your project
2. Install MCP dependencies: task-master, context7, playwright
3. Use: @routing-agent your request

System Management

Using the VAN Maintenance Agent

The van-maintenance-agent manages the agent ecosystem and handles system maintenance:

@van-maintenance-agent [maintenance task]

Common Use Cases:

🆕 Adding New Agents

@van-maintenance-agent integrate new agent [agent-name] into the ecosystem

• Automatically updates interaction diagrams
• Adds agent to categorization system
• Creates proper handoff relationships
• Updates workflow patterns

🔧 Troubleshooting Agent Issues

@van-maintenance-agent fix mermaid syntax errors in agent files
@van-maintenance-agent update agent relationships after changes
@van-maintenance-agent validate handoff token consistency

📊 System Health Checks

@van-maintenance-agent perform comprehensive ecosystem audit
@van-maintenance-agent analyze agent categorization accuracy
@van-maintenance-agent review workflow pattern coverage

🛠️ Documentation Maintenance

@van-maintenance-agent update agent interaction documentation
@van-maintenance-agent fix outdated agent relationships
@van-maintenance-agent validate ecosystem integrity

The van-agent ensures your agent ecosystem stays healthy and properly coordinated as it evolves.

Architecture

Central Hub (routing-agent): Graph-based semantic analysis, JIT context loading Quality Gates: 6 mandatory gates with no bypass (Planning → Infrastructure → Implementation → Testing → Polish → Completion) Specialized Agents: Ultra-narrow domain focus (component, feature, infrastructure, testing, polish) Validation: Progressive retry system (HANDOFF_TOKEN temporarily disabled for redesign)

Hub-and-Spoke Coordination Architecture

graph TB
    subgraph "Central Hub"
        RA[🎯 Routing Agent<br/>Semantic Analysis]
    end
    
    subgraph "Coordination"  
        EPM[🏗️ Project Manager<br/>6 Quality Gates]
    end
    
    subgraph "Research"
        PRD[📋 PRD Research<br/>Context7]
        RESEARCH[🔬 Research Agent<br/>Analysis]
    end
    
    subgraph "Implementation"
        COMP[🎨 Components<br/>UI & Styling]
        FEAT[💾 Features<br/>Business Logic]
        INFRA[🏗️ Infrastructure<br/>Build & Deploy]
        TEST[🧪 Testing<br/>Validation]
        POLISH[✨ Polish<br/>Performance]
    end
    
    subgraph "Quality Gates"
        EQG[🛡️ Quality Gate]
        CG[✅ Complete Gate]
        RG[🎯 Ready Gate]
    end
    
    subgraph "System"
        HOOKS[⚙️ Hooks<br/>Validation]
        TM[📊 TaskMaster<br/>State]
    end
    
    %% Hub connections
    RA -.-> EPM
    RA -.-> PRD
    RA -.-> RESEARCH
    RA -.-> COMP
    RA -.-> FEAT
    RA -.-> INFRA
    RA -.-> TEST
    RA -.-> POLISH
    
    %% Return to hub
    EPM -.-> RA
    PRD -.-> RA
    RESEARCH -.-> RA
    COMP -.-> RA
    FEAT -.-> RA
    INFRA -.-> RA
    TEST -.-> RA
    POLISH -.-> RA
    
    %% Quality flow
    EPM --> EQG --> CG --> RG
    
    %% System coordination
    HOOKS -.-> RA
    TM -.-> EPM
    
    classDef hub fill:#ffd700,stroke:#ff8c00,stroke-width:6px,color:#1a1a1a,font-size:20px
    classDef coordination fill:#e3f2fd,stroke:#1976d2,stroke-width:4px,color:#1a1a1a,font-size:16px
    classDef research fill:#f3e5f5,stroke:#7b1fa2,stroke-width:4px,color:#1a1a1a,font-size:16px
    classDef implementation fill:#e8f5e8,stroke:#2d5a27,stroke-width:3px,color:#1a1a1a,font-size:16px
    classDef quality fill:#fce4ec,stroke:#a91e63,stroke-width:4px,color:#1a1a1a,font-size:16px
    classDef system fill:#fff3e0,stroke:#f57c00,stroke-width:3px,color:#1a1a1a,font-size:16px
    
    class RA hub
    class EPM coordination
    class PRD,RESEARCH research
    class COMP,FEAT,INFRA,TEST,POLISH implementation
    class EQG,CG,RG quality
    class HOOKS,TM system

Loading

Semantic Analysis & JIT Context Loading

graph TD
    REQUEST[👤 REQUEST] --> PARSE[🧠 SEMANTIC<br/>ANALYSIS]
    
    PARSE --> CLASSIFY{🎯 INTENT<br/>CLASSIFICATION}
    
    CLASSIFY -->|UI| UI[🎨 UI CONTEXT<br/>JIT LOAD]
    CLASSIFY -->|DATA| DATA[💾 DATA CONTEXT<br/>JIT LOAD]
    CLASSIFY -->|BUILD| BUILD[🏗️ BUILD CONTEXT<br/>JIT LOAD]
    CLASSIFY -->|TEST| TEST[🧪 TEST CONTEXT<br/>JIT LOAD]
    CLASSIFY -->|INFO| INFO[🔬 RESEARCH CONTEXT<br/>JIT LOAD]
    
    UI --> AGENT_UI[🎨 COMPONENT<br/>AGENT]
    DATA --> AGENT_DATA[💾 FEATURE<br/>AGENT]
    BUILD --> AGENT_BUILD[🏗️ INFRASTRUCTURE<br/>AGENT]
    TEST --> AGENT_TEST[🧪 TESTING<br/>AGENT]
    INFO --> AGENT_INFO[🔬 RESEARCH<br/>AGENT]
    
    AGENT_UI --> TOKEN[🔒 HANDOFF<br/>TOKEN]
    AGENT_DATA --> TOKEN
    AGENT_BUILD --> TOKEN
    AGENT_TEST --> TOKEN
    AGENT_INFO --> TOKEN
    
    TOKEN --> VALIDATE{✅ VALID?}
    
    VALIDATE -->|PASS| SUCCESS[🎯 EXECUTE<br/>WITH CONTEXT]
    VALIDATE -->|FAIL| RETRY[🔄 RETRY<br/>3X MAX]
    
    RETRY --> ESCALATE[🆘 ESCALATE<br/>TO PM]
    
    SUCCESS -.-> HUB[🔄 RETURN<br/>TO HUB]
    ESCALATE -.-> HUB
    HUB -.-> PARSE
    
    classDef input fill:#e1f5fe,stroke:#0277bd,stroke-width:4px,color:#1a1a1a,font-size:18px
    classDef semantic fill:#ffd700,stroke:#ff8c00,stroke-width:6px,color:#1a1a1a,font-size:20px
    classDef classify fill:#e3f2fd,stroke:#1976d2,stroke-width:4px,color:#1a1a1a,font-size:16px
    classDef context fill:#f3e5f5,stroke:#7b1fa2,stroke-width:3px,color:#1a1a1a,font-size:16px
    classDef agent fill:#e8f5e8,stroke:#2d5a27,stroke-width:3px,color:#1a1a1a,font-size:16px
    classDef token fill:#fce4ec,stroke:#a91e63,stroke-width:4px,color:#1a1a1a,font-size:16px
    classDef validate fill:#fff3e0,stroke:#f57c00,stroke-width:3px,color:#1a1a1a,font-size:16px
    classDef system fill:#f1f8e9,stroke:#689f38,stroke-width:3px,color:#1a1a1a,font-size:16px
    
    class REQUEST input
    class PARSE semantic
    class CLASSIFY classify
    class UI,DATA,BUILD,TEST,INFO context
    class AGENT_UI,AGENT_DATA,AGENT_BUILD,AGENT_TEST,AGENT_INFO agent
    class TOKEN token
    class VALIDATE validate
    class SUCCESS,RETRY,ESCALATE,HUB system

Loading

Research Results

Context Engineering Hypothesis Testing:

• Context retention - Hub-and-spoke coordination vs peer communication patterns
• Instruction comprehension - HANDOFF_TOKEN validation vs standard agent handoffs
• Coordination reliability - Progressive retry systems vs single-attempt coordination

Dependencies

claude mcp add task-master -s user -- npx -y --package=task-master-ai task-master-ai
claude mcp add context7 -s user -- npx -y context7-server
claude mcp add playwright -s user -- npx -y playwright-mcp-server

Key Files

• .claude/agents/routing-agent.md - Central hub with graph-based routing
• .claude/agents/enhanced-project-manager-agent.md - 6-gate quality process
• .claude/agents/van-maintenance-agent.md - Agent ecosystem maintenance
• .claude/hooks/routing-executor.sh - Routing validation (HANDOFF_TOKEN disabled)

Built on: cursor-memory-bank (2,400+ stars) research into agent coordination problems.