Ep 12: Birthing a Conversational Bot — Chat Trigger & AI Agent Node Deep Dive

⏱ Est. reading time: 12 min Updated on 4/9/2026

AI Agent Internal Architecture

The AI Agent node isn't a simple "call API once" — it's an agentic decision loop:

graph TB
    subgraph "AI Agent Internal Execution Loop"
        Input[📥 User Input] --> Think[🧠 LLM Thinks]
        Think --> Decision{Need a tool?}
        Decision -->|"No → direct answer"| Answer[📤 Generate Reply]
        Decision -->|"Yes → pick tool"| ToolCall[🔧 Call Tool]
        ToolCall --> ToolResult[📋 Get Result]
        ToolResult --> Think2[🧠 Think Again]
        Think2 --> Decision2{Need more tools?}
        Decision2 -->|"Yes"| ToolCall
        Decision2 -->|"No"| Answer
    end
    style Think fill:#8b5cf6,stroke:#7c3aed,color:#fff
    style ToolCall fill:#ff6d5b,stroke:#e55a4e,color:#fff

AI Agent ≠ single API call. It's a multi-turn inner loop — the LLM might decide it needs weather data, then realize it also needs exchange rates, and finally synthesize everything into one comprehensive answer.

1. Chat Trigger

Chat Trigger provides a built-in web chat UI as the entry point for conversational workflows.

// ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
// Chat Trigger output Item structure
// ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
{
  "json": {
    "chatInput": "What's the weather in Beijing?",
    "sessionId": "sess_abc123",      // Session ID for memory management
    "action": "sendMessage"
  }
}
// ⚠️ Enable authentication in production! Otherwise anyone can chat with your Agent

2. Complete Chatbot Workflow

graph TB
    CT[💬 Chat Trigger] --> Agent[🤖 AI Agent]
    subgraph "Agent Sub-nodes"
        Agent --> Model[🧠 OpenAI gpt-4o]
        Agent --> Mem[💾 Window Buffer Memory]
    end
    style Agent fill:#ff6d5b,stroke:#e55a4e,color:#fff

3. System Prompt Design (4 Layers)

graph TB
    L1["🎭 Layer 1: Role Definition
Who are you? Who do you serve?"]
    L2["📏 Layer 2: Behavior Rules
Length, tone, language"]
    L3["🚫 Layer 3: Boundaries
What's forbidden?"]
    L4["📋 Layer 4: Output Format
JSON? Markdown? Plain text?"]
    L1 --> L2 --> L3 --> L4

4. Multi-Turn Conversation Sequence

sequenceDiagram
    participant User as 👤 User
    participant CT as 💬 Chat Trigger
    participant Agent as 🤖 AI Agent
    participant Mem as 💾 Memory
    participant LLM as 🧠 GPT-4o
    
    Note over User,LLM: === Turn 1 ===
    User->>CT: "Hello, what can you do?"
    CT->>Agent: {chatInput: "Hello...", sessionId: "s1"}
    Agent->>Mem: Read session "s1" history → empty
    Agent->>LLM: [System Prompt] + [User message]
    LLM-->>Agent: "Hi! I'm your AI assistant..."
    Agent->>Mem: Save: [user Q, AI A]
    
    Note over User,LLM: === Turn 2 ===
    User->>CT: "How do I configure Webhooks?"
    Agent->>Mem: Read history → [Turn 1 context]
    Agent->>LLM: [System] + [History] + [New message]
    Note over LLM: Model sees full context
    LLM-->>Agent: "Here are the steps..."

Next Episode

Ep 13 dives deep into Memory node internals — Window Buffer, Token Buffer, and cross-session long-term memory management.