advanced_agent_streaming

Advanced Agent Streaming: 1 Agent + 2 Subagents + 2 Tools

This example demonstrates advanced streaming capabilities with a focused architecture: 1 main agent coordinating 2 specialized subagents using 2 tools. This showcases the Agent SDK's streaming implementation with multi-agent coordination, native thinking support, and reliable final answer delivery.

🎯 Multi-LLM Provider Support

This example now supports OpenAI, Anthropic, and Google Gemini providers:

• Auto-detection based on available API keys
• Configurable models via environment variables
• Native thinking support for compatible models (Anthropic Claude, Gemini 2.5)

Architecture Overview

Project Manager (Main Agent)
├── Tool 1: Market Data Lookup (Mock)
├── Tool 2: Trend Analysis (Mock)
├── Subagent 1: Research Assistant
└── Subagent 2: Data Analyst

🎯 Core Components

1 Main Agent: Project Manager

• Role: Orchestrates the entire workflow with market analysis focus
• Tools: Market Data Lookup + Trend Analysis (simplified mock tools)
• Capabilities: Extended thinking tokens, iterative tool execution, final answer synthesis
• Responsibilities: Task breakdown, data gathering, analysis, and comprehensive reporting

2 Subagents

• Research Assistant: Trend analysis and information organization
• Data Analyst: Market data lookup and statistical analysis

2 Tools

• Market Data Lookup: Provides market statistics (growth rates, market size, segments, regions)
• Trend Analysis: Offers market forecasting insights (global, mobile, social, B2B trends)

🎬 Advanced Streaming Features

• Real-time agent streaming with minimal, clean visualization
• Anthropic Extended Thinking - Gray thinking content, white final answers
• Iterative tool execution with max iterations control
• Final answer synthesis - Always delivers conclusions after tool execution
• Performance metrics and monitoring
• Smart thinking mode detection - Automatic transition from thinking to content

🧠 AI Capabilities Demonstrated

• Extended Thinking Tokens - Real-time reasoning process in gray text
• Tool Integration - Seamless mock tool execution within streaming
• Iterative Tool Calling - Multiple tool calls with result feedback to LLM
• Final Answer Generation - Guaranteed synthesis after tool completion
• Agent Coordination - Multi-agent task delegation and synthesis

Example Scenario

The example runs a comprehensive E-commerce Market Analysis for Q1 2025 Business Planning that demonstrates:

1. Project Coordination - Main agent plans and delegates tasks
2. Information Gathering - Research assistant capabilities
3. Data Analysis - Analyst performs calculations and projections
4. Tool Usage - Calculator and web search integration
5. Synthesis - Main agent combines findings into actionable recommendations

Environment Variables

Required

# LLM Provider (choose one)
export ANTHROPIC_API_KEY="your_anthropic_key"
export OPENAI_API_KEY="your_openai_key"
export GEMINI_API_KEY="your_gemini_key"

# Provider selection (optional, auto-detects based on available API keys)
export LLM_PROVIDER="anthropic"  # or "openai" or "gemini"

# Model selection (optional, uses provider defaults)
export ANTHROPIC_MODEL="claude-3-7-sonnet"  # default
export OPENAI_MODEL="gpt-4o"                # default
export GEMINI_MODEL="gemini-1.5-flash"      # default, or "gemini-2.5-flash" for thinking

Optional (for enhanced functionality)

# Web search capabilities
export SERPER_API_KEY="your_serper_key"

# GitHub repository analysis
export GITHUB_TOKEN="your_github_token"

Usage

Basic Usage

cd examples/advanced_agent_streaming
go run main.go

With Full API Integration

export ANTHROPIC_API_KEY="your_key"
export SERPER_API_KEY="your_serper_key"
export GITHUB_TOKEN="your_github_token"
go run main.go

With OpenAI (Full Support)

export OPENAI_API_KEY="your_key"
export LLM_PROVIDER="openai"
go run main.go

Supported Models:

• gpt-4o - Full streaming, tools, and reasoning support
• gpt-4-turbo - Full streaming and tools support
• o1-mini, o1-preview - Built-in reasoning only (no external tools)

Output Features

🎨 Rich Terminal Visualization

• Color-coded output for different event types
• Progress indicators for long-running operations
• Thinking process visualization with expandable blocks
• Tool execution tracking with timing information
• Performance metrics dashboard

📊 Streaming Metrics

• Real-time event counting
• Content length tracking
• Events per second measurement
• Tool usage statistics
• Error rate monitoring

🔧 Tool Execution Tracking

TOOL EXECUTION: market_data_lookup
├─ Arguments: {"query": "growth_rates"}
├─ Status: executing
└─ Executing...
├─ Result: Market Growth Rates:
│   - E-commerce: 16.5% YoY
│   - Mobile commerce: 23.2% YoY
│   - B2B e-commerce: 18.7% YoY
│   - Social commerce: 34.1% YoY
└─ Duration: <nil>

💭 Thinking Process Display

THINKING BLOCK #1
────────────────────────────────────────────────────────────
Let me analyze this request and execute a plan for comprehensive
e-commerce market analysis.

I need to:
1. Analyze current e-commerce trends and growth patterns
2. Calculate market projections for Q1 2025
3. Identify opportunities/risks for expansion
4. Provide data-backed recommendations

I'll start with trend analysis tool to understand overall trends,
then use market data lookup for specific statistics...
────────────────────────────────────────────────────────────
Final response:
Based on the comprehensive analysis using both trend analysis and
market data tools, here are my findings and recommendations...

Architecture

Agent Hierarchy

Project Manager (Main Agent)
├── Research Assistant
│   └── Trend Analysis Tool
├── Data Analyst
│   └── Market Data Lookup Tool
└── Main Agent Tools
    ├── Market Data Lookup Tool
    └── Trend Analysis Tool

Streaming Flow

LLM (Anthropic/OpenAI)
  ↓ SSE Events
Agent Layer
  ↓ Agent Events
Tool Execution
  ↓ Tool Results
Subagent Delegation
  ↓ Coordinated Results
Terminal Display

Recent Improvements

🔧 Streaming Reliability Fixes

• Final Answer Delivery: Fixed issue where agents would complete tool execution but not display final synthesized responses
• Smart Thinking Mode Detection: Improved transition from thinking (gray) to final content (white)
• Iterative Tool Calling: Proper implementation of tool result feedback to LLM for continued conversation
• Debug Logging: Comprehensive logging for troubleshooting streaming issues

🛠️ Tool System Improvements

• Simplified Mock Tools: Replaced complex calculator with reliable mock data tools
• No External Dependencies: All tools work without API keys for easy testing
• Consistent Tool Responses: Predictable mock data for demonstration purposes

🎨 Minimal Styling

• Clean Output: Removed excessive colors and emojis as requested
• Focus on Content: Gray for thinking, white for answers - minimal distraction
• Professional Appearance: Clean terminal output suitable for production demos

🤖 OpenAI Implementation Completed

• Iterative Tool Calling: Added missing iterative tool loop to GenerateWithToolsStream matching Anthropic implementation
• Final Answer Synthesis: Fixed critical bug where OpenAI streaming would execute tools but never provide final answers
• Tool Result Feedback: Implemented proper tool result feedback to LLM across iterations with maxIterations control
• Parameter Handling: Fixed tool message parameter order (ToolMessage(content, tool_call_id) vs ToolMessage(tool_call_id, content))
• o1 Model Support: Added o1 model compatibility (no system messages, no custom temperature, no tools)
• Multi-Model Support: Both Anthropic and OpenAI streaming now work identically with reasoning and tools

Configuration Options

Streaming Configuration

streamConfig := interfaces.StreamConfig{
    BufferSize:          1000, // Large buffer for complex ops
    IncludeThinking:     true, // Show reasoning process
    IncludeToolProgress: true, // Track tool execution
}

Performance Tuning

• Buffer sizes from 500-1000 for complex operations
• Concurrent tool execution for parallel processing
• Memory optimization for long conversations
• Error recovery with graceful degradation

Mock Mode

When API keys are not provided, the example runs in mock mode with simulated:

• Web search results
• GitHub repository data
• Realistic response timing
• Full streaming visualization

This allows testing the streaming architecture without external dependencies.

Advanced Features

Performance Monitoring

Real-time tracking of:

• Event processing speed
• Tool execution timing
• Memory usage patterns
• Error rates and recovery

Error Handling

• Graceful degradation to mock tools
• Retry mechanisms for transient failures
• Clear error reporting in stream
• Recovery and continuation

Scalability Features

• Large buffer sizes for high-throughput scenarios
• Efficient event processing
• Memory-conscious design
• Concurrent tool execution

Integration Examples

This example demonstrates patterns for:

• Production deployment of streaming agents
• Multi-tool coordination in real applications
• Subagent architecture for complex workflows
• Performance monitoring and optimization
• Error handling and resilience patterns

Dependencies

• Agent SDK Go framework
• LLM providers (Anthropic Claude / OpenAI GPT)
• Web search API (Serper - optional)
• GitHub API (optional)
• Terminal with color support for best experience

Learning Outcomes

After running this example, you'll understand:

• How to implement complex streaming agent architectures
• Multi-agent coordination patterns
• Tool integration with streaming
• Performance monitoring and optimization
• Production-ready error handling
• Advanced terminal visualization techniques