bzzz/old-docs/MCP_INTEGRATION_DESIGN.md

BZZZ v2 MCP Integration Design

GPT-4 Agent Framework for Distributed P2P Collaboration

Executive Summary

This document outlines the comprehensive Model Context Protocol (MCP) integration for BZZZ v2, enabling GPT-4 agents to operate as first-class citizens within the distributed P2P task coordination system. The integration provides a bridge between OpenAI's GPT-4 models and the existing libp2p-based BZZZ infrastructure, creating a hybrid human-AI collaboration environment.

---

1. MCP Server Architecture

1.1 Core MCP Server Design

interface BzzzMcpServer {
  // Protocol Operations
  tools: {
    bzzz_announce: ToolDefinition;
    bzzz_lookup: ToolDefinition;
    bzzz_get: ToolDefinition;
    bzzz_post: ToolDefinition;
    bzzz_thread: ToolDefinition;
    bzzz_subscribe: ToolDefinition;
  };
  
  // Agent Management
  agentLifecycle: AgentLifecycleManager;
  conversationManager: ConversationManager;
  costTracker: OpenAICostTracker;
  
  // BZZZ Protocol Integration
  p2pNode: P2PNodeInterface;
  pubsubManager: PubSubManager;
  hypercoreLogger: HypercoreLogger;
}

1.2 MCP Tool Registry

The MCP server exposes BZZZ protocol operations as standardized tools that GPT-4 agents can invoke:

Core Protocol Tools

1. bzzz_announce - Agent presence announcement

{
  "name": "bzzz_announce",
  "description": "Announce agent presence and capabilities on the BZZZ network",
  "inputSchema": {
    "type": "object",
    "properties": {
      "agent_id": {"type": "string", "description": "Unique agent identifier"},
      "role": {"type": "string", "description": "Agent role (architect, reviewer, etc.)"},
      "capabilities": {"type": "array", "items": {"type": "string"}},
      "specialization": {"type": "string"},
      "max_tasks": {"type": "number", "default": 3}
    }
  }
}

2. bzzz_lookup - Semantic address discovery

{
  "name": "bzzz_lookup",
  "description": "Discover agents and resources using semantic addressing",
  "inputSchema": {
    "type": "object",
    "properties": {
      "semantic_address": {
        "type": "string", 
        "description": "Format: bzzz://agent:role@project:task/path"
      },
      "filter_criteria": {
        "type": "object",
        "properties": {
          "expertise": {"type": "array"},
          "availability": {"type": "boolean"},
          "performance_threshold": {"type": "number"}
        }
      }
    }
  }
}

3. bzzz_get - Content retrieval from addresses

{
  "name": "bzzz_get",
  "description": "Retrieve content from BZZZ semantic addresses",
  "inputSchema": {
    "type": "object",
    "properties": {
      "address": {"type": "string"},
      "include_metadata": {"type": "boolean", "default": true},
      "max_history": {"type": "number", "default": 10}
    }
  }
}

4. bzzz_post - Event/message posting

{
  "name": "bzzz_post",
  "description": "Post events or messages to BZZZ addresses",
  "inputSchema": {
    "type": "object",
    "properties": {
      "target_address": {"type": "string"},
      "message_type": {"type": "string"},
      "content": {"type": "object"},
      "priority": {"type": "string", "enum": ["low", "medium", "high", "urgent"]},
      "thread_id": {"type": "string", "optional": true}
    }
  }
}

5. bzzz_thread - Conversation management

{
  "name": "bzzz_thread",
  "description": "Manage threaded conversations between agents",
  "inputSchema": {
    "type": "object",
    "properties": {
      "action": {"type": "string", "enum": ["create", "join", "leave", "list", "summarize"]},
      "thread_id": {"type": "string", "optional": true},
      "participants": {"type": "array", "items": {"type": "string"}},
      "topic": {"type": "string", "optional": true}
    }
  }
}

6. bzzz_subscribe - Real-time event subscription

{
  "name": "bzzz_subscribe",
  "description": "Subscribe to real-time events from BZZZ network",
  "inputSchema": {
    "type": "object",
    "properties": {
      "event_types": {"type": "array", "items": {"type": "string"}},
      "filter_address": {"type": "string", "optional": true},
      "callback_webhook": {"type": "string", "optional": true}
    }
  }
}

---

2. GPT-4 Agent Framework

2.1 Agent Specialization Definitions

Core Agent Types

1. Architect Agent (bzzz://architect@*)

{
  "role": "architect",
  "capabilities": [
    "system_design",
    "architecture_review",
    "technology_selection",
    "scalability_analysis"
  ],
  "reasoning_prompts": {
    "system": "You are a senior software architect specializing in distributed systems...",
    "task_analysis": "Analyze this task from an architectural perspective...",
    "collaboration": "Coordinate with other architects and provide technical guidance..."
  },
  "interaction_patterns": {
    "peer_architects": "collaborative_review",
    "developers": "guidance_provision",
    "reviewers": "design_validation"
  }
}

2. Code Reviewer Agent (bzzz://reviewer@*)

{
  "role": "reviewer",
  "capabilities": [
    "code_review",
    "security_analysis", 
    "performance_optimization",
    "best_practices_enforcement"
  ],
  "reasoning_prompts": {
    "system": "You are a senior code reviewer focused on quality and security...",
    "review_criteria": "Evaluate code changes against these criteria...",
    "feedback_delivery": "Provide constructive feedback to developers..."
  }
}

3. Documentation Agent (bzzz://docs@*)

{
  "role": "documentation",
  "capabilities": [
    "technical_writing",
    "api_documentation", 
    "user_guides",
    "knowledge_synthesis"
  ],
  "reasoning_prompts": {
    "system": "You specialize in creating clear, comprehensive technical documentation...",
    "content_analysis": "Analyze technical content and identify documentation needs...",
    "audience_adaptation": "Adapt documentation for different audience levels..."
  }
}

2.2 Agent Lifecycle Management

Agent States and Transitions

stateDiagram-v2
    [*] --> Initializing
    Initializing --> Idle: Registration Complete
    Idle --> Active: Task Assigned
    Active --> Collaborating: Multi-agent Context
    Collaborating --> Active: Individual Work
    Active --> Idle: Task Complete
    Idle --> Terminating: Shutdown Signal
    Terminating --> [*]
    
    Active --> Escalating: Human Intervention Needed
    Escalating --> Active: Issue Resolved
    Escalating --> Terminating: Unresolvable Issue

Lifecycle Implementation

type GPTAgent struct {
    ID           string
    Role         AgentRole
    State        AgentState
    Capabilities []string
    
    // OpenAI Configuration
    APIKey       string
    Model        string // gpt-4, gpt-4-turbo, etc.
    TokenLimit   int
    
    // BZZZ Integration
    P2PNode      *p2p.Node
    PubSub       *pubsub.PubSub
    Logger       *logging.HypercoreLog
    
    // Conversation Context
    ActiveThreads map[string]*ConversationThread
    Memory        *AgentMemory
    
    // Cost Management
    TokenUsage   *TokenUsageTracker
    CostLimits   *CostLimitConfig
}

func (agent *GPTAgent) Initialize() error {
    // Register with BZZZ network
    if err := agent.announcePresence(); err != nil {
        return err
    }
    
    // Subscribe to relevant topics
    if err := agent.subscribeToBzzzTopics(); err != nil {
        return err
    }
    
    // Initialize conversation memory
    agent.Memory = NewAgentMemory(agent.ID)
    
    agent.State = AgentStateIdle
    return nil
}

func (agent *GPTAgent) ProcessTask(task *repository.Task) error {
    agent.State = AgentStateActive
    
    // Create conversation context
    context := agent.buildTaskContext(task)
    
    // Check if collaboration is needed
    if agent.shouldCollaborate(task) {
        return agent.initiateCollaboration(task, context)
    }
    
    // Process individually
    return agent.processIndividualTask(task, context)
}

2.3 Context Sharing and Memory Management

Agent Memory System

type AgentMemory struct {
    WorkingMemory    map[string]interface{}  // Current task context
    EpisodicMemory   []ConversationEpisode   // Past interactions
    SemanticMemory   *KnowledgeGraph         // Domain knowledge
    
    // Conversation History
    ThreadMemories   map[string]*ThreadMemory
    
    // Learning and Adaptation
    PerformanceFeedback []FeedbackEntry
    CollaborationHistory []CollaborationEntry
}

type ConversationEpisode struct {
    Timestamp    time.Time
    Participants []string
    Topic        string
    Summary      string
    Outcome      string
    Lessons      []string
}

---

3. Conversation Integration

3.1 Threaded Conversation Architecture

Thread Management System

type ConversationManager struct {
    activeThreads   map[string]*ConversationThread
    threadIndex     *ThreadIndex
    summaryService  *ThreadSummaryService
    escalationRules *EscalationRuleEngine
}

type ConversationThread struct {
    ID           string
    Topic        string
    Participants []AgentParticipant
    Messages     []ThreadMessage
    State        ThreadState
    
    // Context Management
    SharedContext map[string]interface{}
    DecisionLog   []Decision
    
    // Thread Lifecycle
    CreatedAt    time.Time
    LastActivity time.Time
    AutoClose    bool
    CloseAfter   time.Duration
}

type ThreadMessage struct {
    ID        string
    From      string
    Role      AgentRole
    Content   string
    MessageType MessageType
    Timestamp time.Time
    
    // Threading
    ReplyTo   string
    Reactions []MessageReaction
    
    // GPT-4 Specific
    TokenCount int
    Model      string
    Context    *GPTContext
}

3.2 Multi-Agent Collaboration Patterns

Collaborative Review Pattern

func (cm *ConversationManager) InitiateCollaborativeReview(
    task *repository.Task,
    requiredRoles []AgentRole,
) (*ConversationThread, error) {
    
    // Create thread for collaborative review
    thread := &ConversationThread{
        ID:    generateThreadID("review", task.Number),
        Topic: fmt.Sprintf("Collaborative Review: %s", task.Title),
        State: ThreadStateActive,
    }
    
    // Invite relevant agents
    for _, role := range requiredRoles {
        agents := cm.findAvailableAgents(role)
        for _, agent := range agents[:min(2, len(agents))] {
            thread.Participants = append(thread.Participants, AgentParticipant{
                AgentID: agent.ID,
                Role:    role,
                Status:  ParticipantStatusInvited,
            })
        }
    }
    
    // Set initial context
    thread.SharedContext = map[string]interface{}{
        "task_details": task,
        "review_criteria": getReviewCriteria(task),
        "deadline": calculateReviewDeadline(task),
    }
    
    // Start the conversation
    initialPrompt := cm.buildCollaborativeReviewPrompt(task, thread)
    if err := cm.postInitialMessage(thread, initialPrompt); err != nil {
        return nil, err
    }
    
    return thread, nil
}

Escalation Workflow Pattern

type EscalationRuleEngine struct {
    rules []EscalationRule
}

type EscalationRule struct {
    Name        string
    Conditions  []EscalationCondition
    Actions     []EscalationAction
    Priority    int
}

type EscalationCondition struct {
    Type       string // "thread_duration", "consensus_failure", "error_rate"
    Threshold  interface{}
    Timeframe  time.Duration
}

func (ere *EscalationRuleEngine) CheckEscalation(thread *ConversationThread) []EscalationAction {
    var actions []EscalationAction
    
    for _, rule := range ere.rules {
        if ere.evaluateConditions(rule.Conditions, thread) {
            actions = append(actions, rule.Actions...)
        }
    }
    
    return actions
}

// Example escalation scenarios
var DefaultEscalationRules = []EscalationRule{
    {
        Name: "Long Running Thread",
        Conditions: []EscalationCondition{
            {Type: "thread_duration", Threshold: 2 * time.Hour, Timeframe: 0},
            {Type: "no_progress", Threshold: true, Timeframe: 30 * time.Minute},
        },
        Actions: []EscalationAction{
            {Type: "notify_human", Target: "project_manager"},
            {Type: "request_expert", Expertise: []string{"domain_expert"}},
        },
    },
    {
        Name: "Consensus Failure",
        Conditions: []EscalationCondition{
            {Type: "disagreement_count", Threshold: 3, Timeframe: 0},
            {Type: "no_resolution", Threshold: true, Timeframe: 1 * time.Hour},
        },
        Actions: []EscalationAction{
            {Type: "escalate_to_architect", Priority: "high"},
            {Type: "create_decision_thread", Participants: []string{"senior_architect"}},
        },
    },
}

---

4. CHORUS Integration Patterns

4.1 SLURP Context Integration

SLURP Event Generation from HMMM Consensus

type SLURPIntegrationService struct {
    slurpClient     *slurp.Client
    conversationMgr *ConversationManager
    eventGenerator  *ConsensusEventGenerator
}

func (sis *SLURPIntegrationService) GenerateSLURPEventFromConsensus(
    thread *ConversationThread,
    consensus *ThreadConsensus,
) (*slurp.ContextEvent, error) {
    
    // Analyze conversation for insights
    insights := sis.extractInsights(thread)
    
    // Generate structured event
    event := &slurp.ContextEvent{
        Type:        "agent_consensus",
        Source:      "bzzz_mcp_integration", 
        Timestamp:   time.Now(),
        Context: slurp.ContextData{
            ConversationID: thread.ID,
            Participants:   getParticipantRoles(thread.Participants),
            Topic:          thread.Topic,
            Insights:       insights,
            DecisionPoints: consensus.Decisions,
            Confidence:     consensus.ConfidenceScore,
        },
        Metadata: map[string]interface{}{
            "thread_duration": thread.LastActivity.Sub(thread.CreatedAt).Minutes(),
            "message_count":   len(thread.Messages),
            "agent_count":     len(thread.Participants),
            "consensus_type":  consensus.Type,
        },
    }
    
    // Send to SLURP system
    if err := sis.slurpClient.SubmitContextEvent(event); err != nil {
        return nil, fmt.Errorf("failed to submit SLURP event: %w", err)
    }
    
    // Notify BZZZ network of event generation
    sis.notifyEventGenerated(thread, event)
    
    return event, nil
}

4.2 WHOOSH Orchestration Integration

GPT-4 Agent Registration with WHOOSH

type WHOOSHIntegrationService struct {
    whooshClient *whoosh.Client
    agentRegistry map[string]*GPTAgent
}

func (wis *WHOOSHIntegrationService) RegisterGPTAgentWithWHOOSH(
    agent *GPTAgent,
) error {
    
    // Create WHOOSH agent registration
    registration := &whoosh.AgentRegistration{
        AgentID:      agent.ID,
        Type:        "gpt_agent",
        Role:        string(agent.Role),
        Capabilities: agent.Capabilities,
        Metadata: map[string]interface{}{
            "model":           agent.Model,
            "max_tokens":      agent.TokenLimit,
            "cost_per_token":  getTokenCost(agent.Model),
            "bzzz_address":    fmt.Sprintf("bzzz://%s:%s@*", agent.ID, agent.Role),
        },
        Endpoints: whoosh.AgentEndpoints{
            StatusCheck: fmt.Sprintf("http://mcp-server:8080/agents/%s/status", agent.ID),
            TaskAssign:  fmt.Sprintf("http://mcp-server:8080/agents/%s/tasks", agent.ID),
            Collaborate: fmt.Sprintf("http://mcp-server:8080/agents/%s/collaborate", agent.ID),
        },
        HealthCheck: whoosh.HealthCheckConfig{
            Interval:    30 * time.Second,
            Timeout:     10 * time.Second,
            Retries:     3,
        },
    }
    
    // Submit registration
    if err := wis.whooshClient.RegisterAgent(registration); err != nil {
        return fmt.Errorf("failed to register with WHOOSH: %w", err)
    }
    
    // Start health reporting
    go wis.reportAgentHealth(agent)
    
    return nil
}

4.3 TGN (The Garden Network) Connectivity

Cross-Network Agent Discovery

type TGNConnector struct {
    tgnClient     *tgn.Client
    bzzzNetwork   *BzzzNetwork
    agentRegistry *AgentRegistry
}

func (tgn *TGNConnector) DiscoverCrossNetworkAgents(
    query *AgentDiscoveryQuery,
) ([]*RemoteAgent, error) {
    
    // Query TGN for agents matching criteria
    tgnQuery := &tgn.AgentQuery{
        Capabilities: query.RequiredCapabilities,
        Role:        query.Role,
        Network:     "bzzz",
        Available:   true,
    }
    
    remoteAgents, err := tgn.tgnClient.DiscoverAgents(tgnQuery)
    if err != nil {
        return nil, err
    }
    
    // Convert TGN agents to BZZZ addressable agents
    var bzzzAgents []*RemoteAgent
    for _, remote := range remoteAgents {
        bzzzAgent := &RemoteAgent{
            ID:           remote.ID,
            Network:      remote.Network,
            BzzzAddress: fmt.Sprintf("bzzz://%s:%s@%s/*", 
                        remote.ID, remote.Role, remote.Network),
            Capabilities: remote.Capabilities,
            Endpoint:    remote.Endpoint,
        }
        bzzzAgents = append(bzzzAgents, bzzzAgent)
    }
    
    return bzzzAgents, nil
}

---

5. Implementation Roadmap

5.1 Phase 1: Core MCP Infrastructure (Weeks 1-2)

Week 1: MCP Server Foundation

• Implement basic MCP server with tool registry
• Create OpenAI API integration wrapper
• Establish P2P node connection interface
• Basic agent lifecycle management

Key Deliverables:

• MCP server binary with basic tool definitions
• OpenAI GPT-4 integration module
• Agent registration and deregistration flows

Week 2: Protocol Tool Implementation

• Implement all six core bzzz:// protocol tools
• Add semantic addressing support
• Create pubsub message routing
• Basic conversation threading

Key Deliverables:

• Full protocol tool suite
• Address resolution system
• Message routing infrastructure

5.2 Phase 2: Agent Framework (Weeks 3-4)

Week 3: Agent Specializations

• Define role-based agent templates
• Implement reasoning prompt systems
• Create capability matching logic
• Agent memory management

Week 4: Collaboration Patterns

• Multi-agent conversation threading
• Consensus building algorithms
• Escalation rule engine
• Human intervention workflows

5.3 Phase 3: CHORUS Integration (Weeks 5-6)

Week 5: SLURP Integration

• Consensus-to-SLURP event generation
• Context relevance scoring
• Feedback loop implementation
• Performance optimization

Week 6: WHOOSH & TGN Integration

• Agent registration with WHOOSH
• Cross-network agent discovery
• Task orchestration bridging
• Network topology management

5.4 Phase 4: Production Readiness (Weeks 7-8)

Week 7: Monitoring & Cost Management

• OpenAI cost tracking and limits
• Performance monitoring dashboards
• Conversation analytics
• Agent efficiency metrics

Week 8: Testing & Deployment

• End-to-end integration testing
• Load testing with multiple agents
• Security auditing
• Production deployment automation

---

6. Technical Requirements

6.1 Infrastructure Requirements

Server Specifications

• CPU: 8+ cores for concurrent agent processing
• RAM: 32GB+ for conversation context management
• Storage: 1TB+ SSD for conversation history and logs
• Network: High-speed connection for P2P communication

Software Dependencies

• Go 1.21+: For BZZZ P2P integration
• Node.js 18+: For MCP server implementation
• Docker: For containerized deployment
• PostgreSQL 14+: For conversation persistence

6.2 Security Considerations

API Key Management

• OpenAI API keys stored in secure vault
• Per-agent API key rotation
• Usage monitoring and alerting
• Rate limiting and quotas

P2P Security

• Message signing and verification
• Agent authentication protocols
• Network access controls
• Audit logging

6.3 Cost Management

Token Usage Optimization

type CostOptimizer struct {
    tokenBudgets    map[string]*TokenBudget
    usageTracking   *UsageTracker
    costCalculator  *CostCalculator
}

func (co *CostOptimizer) OptimizeConversation(thread *ConversationThread) {
    // Compress context when approaching limits
    if thread.EstimatedTokens() > thread.TokenBudget * 0.8 {
        co.compressConversationHistory(thread)
    }
    
    // Use cheaper models for routine tasks
    if thread.Complexity < ComplexityThreshold {
        co.assignModel(thread, "gpt-4o-mini")
    }
    
    // Implement conversation summarization
    if len(thread.Messages) > MaxMessagesBeforeSummary {
        co.summarizeAndTruncate(thread)
    }
}

---

7. Code Examples

7.1 MCP Server Implementation

// pkg/mcp/server.go
package mcp

import (
    "context"
    "encoding/json"
    "fmt"
    "net/http"
    
    "github.com/anthonyrawlins/bzzz/p2p"
    "github.com/anthonyrawlins/bzzz/pubsub"
    openai "github.com/sashabaranov/go-openai"
)

type McpServer struct {
    p2pNode      *p2p.Node
    pubsub       *pubsub.PubSub
    openaiClient *openai.Client
    agents       map[string]*GPTAgent
    tools        map[string]ToolHandler
}

func NewMcpServer(apiKey string, node *p2p.Node, ps *pubsub.PubSub) *McpServer {
    server := &McpServer{
        p2pNode:      node,
        pubsub:       ps,
        openaiClient: openai.NewClient(apiKey),
        agents:       make(map[string]*GPTAgent),
        tools:        make(map[string]ToolHandler),
    }
    
    // Register protocol tools
    server.registerProtocolTools()
    
    return server
}

func (s *McpServer) registerProtocolTools() {
    s.tools["bzzz_announce"] = s.handleBzzzAnnounce
    s.tools["bzzz_lookup"] = s.handleBzzzLookup
    s.tools["bzzz_get"] = s.handleBzzzGet
    s.tools["bzzz_post"] = s.handleBzzzPost
    s.tools["bzzz_thread"] = s.handleBzzzThread
    s.tools["bzzz_subscribe"] = s.handleBzzzSubscribe
}

func (s *McpServer) handleBzzzAnnounce(params map[string]interface{}) (interface{}, error) {
    agentID, ok := params["agent_id"].(string)
    if !ok {
        return nil, fmt.Errorf("agent_id is required")
    }
    
    role, ok := params["role"].(string)
    if !ok {
        return nil, fmt.Errorf("role is required")
    }
    
    // Create announcement message
    announcement := map[string]interface{}{
        "agent_id":      agentID,
        "role":         role,
        "capabilities": params["capabilities"],
        "specialization": params["specialization"],
        "max_tasks":    params["max_tasks"],
        "announced_at": time.Now(),
    }
    
    // Publish to BZZZ network
    err := s.pubsub.PublishBzzzMessage(pubsub.CapabilityBcast, announcement)
    if err != nil {
        return nil, fmt.Errorf("failed to announce: %w", err)
    }
    
    return map[string]interface{}{
        "status": "announced",
        "message": fmt.Sprintf("Agent %s (%s) announced to network", agentID, role),
    }, nil
}

func (s *McpServer) handleBzzzLookup(params map[string]interface{}) (interface{}, error) {
    address, ok := params["semantic_address"].(string)
    if !ok {
        return nil, fmt.Errorf("semantic_address is required")
    }
    
    // Parse semantic address (bzzz://agent:role@project:task/path)
    parsedAddr, err := parseSemanticAddress(address)
    if err != nil {
        return nil, fmt.Errorf("invalid semantic address: %w", err)
    }
    
    // Discover matching agents
    agents := s.discoverAgents(parsedAddr, params["filter_criteria"])
    
    return map[string]interface{}{
        "address": address,
        "matches": agents,
        "count":   len(agents),
    }, nil
}

7.2 GPT-4 Agent Implementation

// pkg/agents/gpt_agent.go
package agents

import (
    "context"
    "fmt"
    
    openai "github.com/sashabaranov/go-openai"
    "github.com/anthonyrawlins/bzzz/pubsub"
)

type GPTAgent struct {
    ID           string
    Role         AgentRole
    Model        string
    Client       *openai.Client
    
    SystemPrompt string
    Memory       *AgentMemory
    CostTracker  *CostTracker
    
    // BZZZ Integration
    PubSub       *pubsub.PubSub
    Logger       *logging.HypercoreLog
}

func (agent *GPTAgent) ProcessCollaborativeTask(
    task *repository.Task,
    thread *ConversationThread,
) error {
    
    // Build context from conversation history
    context := agent.buildTaskContext(task, thread)
    
    // Create GPT-4 request
    messages := []openai.ChatCompletionMessage{
        {
            Role:    openai.ChatMessageRoleSystem,
            Content: agent.buildSystemPrompt(task, thread),
        },
    }
    
    // Add conversation history
    for _, msg := range thread.Messages {
        messages = append(messages, openai.ChatCompletionMessage{
            Role:    openai.ChatMessageRoleUser,
            Content: fmt.Sprintf("[%s]: %s", msg.From, msg.Content),
        })
    }
    
    // Add current task context
    messages = append(messages, openai.ChatCompletionMessage{
        Role:    openai.ChatMessageRoleUser,
        Content: agent.formatTaskForGPT(task),
    })
    
    // Make GPT-4 request
    resp, err := agent.Client.CreateChatCompletion(
        context.Background(),
        openai.ChatCompletionRequest{
            Model:     agent.Model,
            Messages:  messages,
            MaxTokens: 2000,
            Tools:     agent.getAvailableTools(),
        },
    )
    
    if err != nil {
        return fmt.Errorf("GPT-4 request failed: %w", err)
    }
    
    // Process response and tool calls
    return agent.processGPTResponse(resp, thread)
}

func (agent *GPTAgent) buildSystemPrompt(task *repository.Task, thread *ConversationThread) string {
    basePrompt := agent.SystemPrompt
    
    // Add role-specific context
    roleContext := agent.getRoleSpecificContext(task)
    
    // Add collaboration context
    collabContext := fmt.Sprintf(
        "\nYou are collaborating with %d other agents in thread '%s'.\n"+
        "Current participants: %s\n"+
        "Thread topic: %s\n"+
        "Your role in this collaboration: %s\n",
        len(thread.Participants)-1,
        thread.ID,
        getParticipantList(thread.Participants),
        thread.Topic,
        agent.Role,
    )
    
    // Add available tools context
    toolsContext := "\nAvailable BZZZ tools:\n"
    for toolName, tool := range agent.getAvailableTools() {
        toolsContext += fmt.Sprintf("- %s: %s\n", toolName, tool.Function.Description)
    }
    
    return basePrompt + roleContext + collabContext + toolsContext
}

7.3 Conversation Threading

// pkg/conversations/thread_manager.go
package conversations

type ThreadManager struct {
    threads       map[string]*ConversationThread
    participants  map[string][]string // agentID -> threadIDs
    summaryEngine *SummaryEngine
    escalationMgr *EscalationManager
}

func (tm *ThreadManager) CreateCollaborativeThread(
    topic string,
    task *repository.Task,
    requiredRoles []AgentRole,
) (*ConversationThread, error) {
    
    thread := &ConversationThread{
        ID:           generateThreadID(topic, task.Number),
        Topic:        topic,
        State:        ThreadStateActive,
        CreatedAt:    time.Now(),
        SharedContext: map[string]interface{}{
            "task":          task,
            "required_roles": requiredRoles,
        },
    }
    
    // Find and invite agents
    for _, role := range requiredRoles {
        agents := tm.findAvailableAgentsByRole(role)
        if len(agents) == 0 {
            return nil, fmt.Errorf("no available agents for role: %s", role)
        }
        
        // Select best agent for this role
        selectedAgent := tm.selectBestAgent(agents, task)
        
        thread.Participants = append(thread.Participants, AgentParticipant{
            AgentID: selectedAgent.ID,
            Role:    role,
            Status:  ParticipantStatusInvited,
        })
    }
    
    // Initialize thread
    tm.threads[thread.ID] = thread
    
    // Send invitations
    for _, participant := range thread.Participants {
        if err := tm.inviteToThread(participant.AgentID, thread); err != nil {
            fmt.Printf("Failed to invite agent %s: %v\n", participant.AgentID, err)
        }
    }
    
    // Start thread monitoring
    go tm.monitorThread(thread)
    
    return thread, nil
}

func (tm *ThreadManager) PostMessage(
    threadID string,
    fromAgent string,
    content string,
    messageType MessageType,
) error {
    
    thread, exists := tm.threads[threadID]
    if !exists {
        return fmt.Errorf("thread %s not found", threadID)
    }
    
    message := ThreadMessage{
        ID:        generateMessageID(),
        From:      fromAgent,
        Content:   content,
        Type:      messageType,
        Timestamp: time.Now(),
    }
    
    thread.Messages = append(thread.Messages, message)
    thread.LastActivity = time.Now()
    
    // Notify all participants
    for _, participant := range thread.Participants {
        if participant.AgentID != fromAgent {
            if err := tm.notifyParticipant(participant.AgentID, thread, message); err != nil {
                fmt.Printf("Failed to notify %s: %v\n", participant.AgentID, err)
            }
        }
    }
    
    // Check for escalation conditions
    if actions := tm.escalationMgr.CheckEscalation(thread); len(actions) > 0 {
        tm.executeEscalationActions(thread, actions)
    }
    
    return nil
}

---

8. Success Metrics

8.1 Performance Metrics

• Agent Response Time: < 30 seconds for routine tasks
• Collaboration Efficiency: 40% reduction in task completion time
• Consensus Success Rate: > 85% of collaborative discussions reach consensus
• Escalation Rate: < 15% of threads require human intervention

8.2 Cost Metrics

• Token Efficiency: < $0.50 per task for routine tasks
• Model Selection Accuracy: > 90% appropriate model selection
• Context Compression Ratio: 70% reduction in token usage through compression

8.3 Quality Metrics

• Code Review Accuracy: > 95% critical issues detected
• Documentation Completeness: > 90% coverage of technical requirements
• Architecture Consistency: > 95% adherence to established patterns

---

9. Security and Compliance

9.1 Data Protection

• All conversation data encrypted at rest and in transit
• Agent memory isolation between different projects
• Automatic PII detection and redaction
• Configurable data retention policies

9.2 Access Control

• Role-based access to different agent capabilities
• Project-level agent permissions
• API key scoping and rotation
• Audit logging of all agent actions

9.3 Compliance Considerations

• GDPR compliance for European operations
• SOC 2 Type II compliance framework
• Regular security audits and penetration testing
• Incident response procedures for AI agent failures

---

This comprehensive design provides the foundation for implementing GPT-4 agents as first-class citizens in the BZZZ v2 distributed system, enabling sophisticated multi-agent collaboration while maintaining the security, performance, and cost-effectiveness required for production deployment.