kailash-kaizen 0.7.1

Advanced AI agent framework built on Kailash SDK

Kaizen - Signature-Based AI Agent Framework

Production-ready AI agents with multi-modal processing, multi-agent coordination, and enterprise features built on Kailash SDK

Kaizen provides a unified BaseAgent architecture where you extend agents for specific use cases, define type-safe Signatures for inputs/outputs, and leverage automatic optimization, error handling, and audit trails.

🎯 What is Kaizen?

Kaizen transforms AI agent development through signature-based programming and a unified BaseAgent architecture. Instead of reinventing agent patterns, extend BaseAgent with domain-specific logic while inheriting production-grade features automatically.

Core Value Propositions

Traditional AI Agent Development:

# Build everything from scratch
class MyAgent:
    def __init__(self, model, temperature, ...):
        self.model = model  # Manual setup
        self.temperature = temperature
        self.memory = []  # Manual memory management
        # ... dozens of lines for error handling, logging, etc.

    def process(self, input_data):
        # Manual prompt construction, error handling, retry logic...
        pass

Kaizen Signature-Based Development:

from kaizen.core.base_agent import BaseAgent
from kaizen.signatures import Signature, InputField, OutputField
from dataclasses import dataclass

# 1. Define configuration
@dataclass
class MyConfig:
    llm_provider: str = "openai"
    model: str = "gpt-4"
    temperature: float = 0.7

# 2. Define signature (type-safe I/O)
class MySignature(Signature):
    question: str = InputField(desc="User question")
    answer: str = OutputField(desc="Agent answer")

# 3. Extend BaseAgent (87% less code, production-ready)
class MyAgent(BaseAgent):
    def __init__(self, config: MyConfig):
        super().__init__(config=config, signature=MySignature())

    def ask(self, question: str):
        return self.run(question=question)  # Auto: logging, error handling, performance tracking

Key Benefits:

• Unified Architecture: BaseAgent provides common infrastructure (87% code reduction)
• Type-Safe Signatures: Define inputs/outputs, framework handles validation
• Auto-Optimization: Automatic async execution, lazy initialization, performance tracking
• Enterprise Ready: Built-in error handling, logging, audit trails, memory management
• Multi-Modal: Vision (Ollama + OpenAI GPT-4V), Audio (Whisper)
• Multi-Agent: Google A2A protocol for semantic capability matching
• Autonomous Tool Calling (v0.7.0): 12 builtin tools with approval workflows
• Bidirectional Control Protocol (v0.7.0): Agent ↔ client communication (questions, approvals, progress)
• Core SDK Compatible: Seamless integration with Kailash workflows

🚀 Quick Start

Installation

# Install Kaizen framework (latest v0.7.0)
pip install kailash-kaizen

# Or specific version
pip install kailash-kaizen==0.7.0

Your First Agent (3 Steps)

from kaizen.agents import SimpleQAAgent
from kaizen.agents.specialized.simple_qa import SimpleQAConfig
from dotenv import load_dotenv

# Load API keys from .env
load_dotenv()

# 1. Create config
config = SimpleQAConfig(
    llm_provider="openai",
    model="gpt-4"
)

# 2. Create agent
agent = SimpleQAAgent(config)

# 3. Execute
result = agent.ask("What is quantum computing?")
print(result["answer"])
print(f"Confidence: {result['confidence']}")

Production Agent with Memory

from kaizen.agents import SimpleQAAgent
from kaizen.agents.specialized.simple_qa import SimpleQAConfig

# Enable memory with max_turns parameter
config = SimpleQAConfig(
    llm_provider="openai",
    model="gpt-4",
    max_turns=10  # Enable BufferMemory
)

agent = SimpleQAAgent(config)

# Use session_id for memory continuity
result1 = agent.ask("My name is Alice", session_id="user123")
result2 = agent.ask("What's my name?", session_id="user123")
print(result2["answer"])  # "Your name is Alice"

🏗️ BaseAgent Architecture

The BaseAgent Pattern

All Kaizen agents follow the same unified architecture:

from kaizen.core.base_agent import BaseAgent
from kaizen.signatures import Signature, InputField, OutputField
from dataclasses import dataclass

# Step 1: Define Configuration (auto-extracted by BaseAgent)
@dataclass
class SimpleQAConfig:
    llm_provider: str = "openai"
    model: str = "gpt-4"
    temperature: float = 0.1
    max_tokens: int = 300
    # BaseAgent auto-extracts: llm_provider, model, temperature, max_tokens, provider_config

# Step 2: Define Signature (type-safe inputs/outputs)
class QASignature(Signature):
    """Answer questions accurately and concisely with confidence scoring."""
    question: str = InputField(desc="The question to answer")
    context: str = InputField(desc="Additional context if available", default="")

    answer: str = OutputField(desc="Clear, accurate answer")
    confidence: float = OutputField(desc="Confidence score 0.0-1.0")
    reasoning: str = OutputField(desc="Brief explanation of reasoning")

# Step 3: Extend BaseAgent
class SimpleQAAgent(BaseAgent):
    """Simple Q&A Agent using BaseAgent architecture."""

    def __init__(self, config: SimpleQAConfig):
        # BaseAgent auto-converts config → BaseAgentConfig
        super().__init__(config=config, signature=QASignature())
        self.qa_config = config

    def ask(self, question: str, context: str = "") -> dict:
        """
        Process a question and return structured answer.

        BaseAgent.run() provides:
        - Automatic logging (LoggingMixin)
        - Performance tracking (PerformanceMixin)
        - Error handling (ErrorHandlingMixin)
        - Memory management (if configured)
        """
        return self.run(question=question, context=context)

What BaseAgent Provides

Automatic Features (inherited by all agents):

1. Config Auto-Extraction: Converts domain config → BaseAgentConfig
2. Async Execution: AsyncSingleShotStrategy for 2-3x performance improvement
3. Error Handling: Automatic retries, timeouts, graceful degradation
4. Performance Tracking: Built-in timing, token counting, cost tracking
5. Structured Logging: Comprehensive logging with context
6. Memory Management: Optional BufferMemory with session support
7. A2A Integration: Auto-generates Agent-to-Agent capability cards
8. Workflow Generation: to_workflow() for Core SDK integration

Code Reduction:

• Traditional agent: ~496 lines
• BaseAgent-based: ~65 lines
• 87% reduction with more features

📚 Available Specialized Agents

Implemented and Production-Ready

from kaizen.agents import (
    # Single-Agent Patterns
    SimpleQAAgent,           # Question answering with confidence scoring
    ChainOfThoughtAgent,     # Step-by-step reasoning
    ReActAgent,              # Reasoning + action cycles
    RAGResearchAgent,        # Research with retrieval-augmented generation
    CodeGenerationAgent,     # Code generation and explanation
    MemoryAgent,             # Memory-enhanced conversations

    # Multi-Modal Agents
    VisionAgent,             # Image analysis (Ollama llava/bakllava + OpenAI GPT-4V)
    TranscriptionAgent,      # Audio transcription (Whisper)
)

Usage Examples

SimpleQAAgent - Question Answering:

from kaizen.agents import SimpleQAAgent
from kaizen.agents.specialized.simple_qa import SimpleQAConfig

config = SimpleQAConfig(llm_provider="openai", model="gpt-4")
agent = SimpleQAAgent(config)
result = agent.ask("What is the capital of France?")
print(result["answer"])  # "Paris"
print(result["confidence"])  # 0.95

ChainOfThoughtAgent - Step-by-Step Reasoning:

from kaizen.agents import ChainOfThoughtAgent
from kaizen.agents.specialized.chain_of_thought import ChainOfThoughtConfig

config = ChainOfThoughtConfig(llm_provider="openai", model="gpt-4")
agent = ChainOfThoughtAgent(config)
result = agent.think("If John has 3 apples and Mary gives him 5 more, how many does he have?")
print(result["reasoning_steps"])  # ["Step 1: John starts with 3 apples", ...]
print(result["final_answer"])     # "8 apples"

VisionAgent - Image Analysis:

from kaizen.agents import VisionAgent, VisionAgentConfig

# Ollama vision (free, local)
config = VisionAgentConfig(llm_provider="ollama", model="bakllava")
agent = VisionAgent(config=config)

result = agent.analyze(
    image="/path/to/receipt.jpg",
    question="What is the total amount?"
)
print(result['answer'])  # "$42.99"

TranscriptionAgent - Audio Transcription:

from kaizen.agents import TranscriptionAgent, TranscriptionAgentConfig

config = TranscriptionAgentConfig()  # Uses Whisper by default
agent = TranscriptionAgent(config=config)

result = agent.transcribe(audio_path="/path/to/audio.mp3")
print(result['transcription'])  # Full text transcription

🎯 Multi-Modal Processing

Vision Processing (Ollama + OpenAI)

from kaizen.agents import VisionAgent, VisionAgentConfig

# Option 1: Ollama (free, local, requires Ollama installed)
ollama_config = VisionAgentConfig(
    llm_provider="ollama",
    model="bakllava"  # or "llava"
)
ollama_agent = VisionAgent(config=ollama_config)

# Option 2: OpenAI GPT-4V (paid API, higher quality)
openai_config = VisionAgentConfig(
    llm_provider="openai",
    model="gpt-4o"
)
openai_agent = VisionAgent(config=openai_config)

# Analyze image
result = ollama_agent.analyze(
    image="/path/to/invoice.jpg",
    question="Extract all line items and totals"
)
print(result['answer'])

Audio Processing (Whisper)

from kaizen.agents import TranscriptionAgent, TranscriptionAgentConfig

config = TranscriptionAgentConfig()
agent = TranscriptionAgent(config=config)

# Transcribe audio file
result = agent.transcribe(audio_path="/path/to/meeting.mp3")
print(result['transcription'])
print(result['duration'])
print(result['language'])

Common Pitfalls - Multi-Modal API

# ❌ WRONG: Using 'prompt' instead of 'question'
result = vision_agent.analyze(image=img, prompt="What is this?")

# ❌ WRONG: Using 'response' key instead of 'answer'
answer = result['response']

# ❌ WRONG: Passing base64 string instead of file path
result = vision_agent.analyze(image=base64_string, question="...")

# ✅ CORRECT: Use 'question' parameter and 'answer' key
result = vision_agent.analyze(image="/path/to/image.png", question="What is this?")
answer = result['answer']

🤝 Multi-Agent Coordination

Google A2A Protocol Integration

Kaizen implements the Google Agent-to-Agent (A2A) protocol for semantic capability matching. No hardcoded if/else logic - agents automatically match tasks based on semantic similarity.

from kaizen.orchestration.patterns.supervisor_worker import SupervisorWorkerPattern
from kaizen.agents import SimpleQAAgent, CodeGenerationAgent, RAGResearchAgent

# Create specialized worker agents
qa_agent = SimpleQAAgent(config=SimpleQAConfig())
code_agent = CodeGenerationAgent(config=CodeConfig())
research_agent = RAGResearchAgent(config=RAGConfig())

# Create pattern with automatic A2A capability matching
pattern = SupervisorWorkerPattern(
    supervisor=supervisor_agent,
    workers=[qa_agent, code_agent, research_agent],
    coordinator=coordinator,
    shared_pool=shared_memory_pool
)

# Semantic task routing (eliminates 40-50% of manual selection logic)
result = pattern.execute_task("Analyze this codebase and suggest improvements")

# A2A automatically selects best worker based on semantic similarity
best_worker = pattern.supervisor.select_worker_for_task(
    task="Analyze sales data and create visualization",
    available_workers=[qa_agent, code_agent, research_agent],
    return_score=True
)
# Returns: {"worker": <RAGResearchAgent>, "score": 0.87}

Available Coordination Patterns

1. SupervisorWorkerPattern - Task delegation with semantic matching (14/14 tests ✅)
2. ConsensusPattern - Group decision-making
3. DebatePattern - Adversarial reasoning
4. SequentialPattern - Step-by-step processing
5. HandoffPattern - Dynamic agent handoff

🛠️ Autonomous Tool Calling (v0.7.0)

Overview

BaseAgent now supports autonomous tool calling with built-in safety controls and approval workflows via MCP (Model Context Protocol). Agents can discover, execute, and chain tools to accomplish complex tasks.

from kaizen.core.base_agent import BaseAgent

# Tools are auto-configured via MCP
agent = BaseAgent(
    config=config,
    signature=signature,
    tools="all"  # Enables 12 builtin tools via MCP
)

# OR configure custom MCP servers:
mcp_servers = [{
    "name": "kaizen_builtin",
    "command": "python",
    "args": ["-m", "kaizen.mcp.builtin_server"],
    "transport": "stdio"
}]
agent = BaseAgent(
    config=config,
    signature=signature,
    custom_mcp_servers=mcp_servers
)

12 Builtin Tools

File Operations (5 tools):

• read_file, write_file, delete_file, list_directory, file_exists

HTTP Requests (4 tools):

• http_get, http_post, http_put, http_delete

System Operations (1 tool):

• bash_command

Web Scraping (2 tools):

• fetch_url, extract_links

Tool Discovery and Execution

# Discover available tools
tools = await agent.discover_tools(category="file", safe_only=True)

# Execute single tool (with approval workflow)
result = await agent.execute_tool(
    tool_name="read_file",
    params={"path": "/tmp/data.txt"}
)

if result.success and result.approved:
    print(f"Content: {result.result['content']}")

# Chain multiple tools
results = await agent.execute_tool_chain([
    {"tool_name": "read_file", "params": {"path": "input.txt"}},
    {"tool_name": "bash_command", "params": {"command": "wc -l input.txt"}},
    {"tool_name": "write_file", "params": {"path": "output.txt", "content": "..."}}
])

Approval Workflows

Tools are classified by danger level:

• SAFE: Auto-approved (no side effects) - list_directory, file_exists
• LOW: Read-only operations - read_file, http_get
• MEDIUM: Data modification - write_file, http_post
• HIGH: Destructive operations - delete_file, bash_command

Non-SAFE tools require explicit approval via the Control Protocol.

🔄 Control Protocol (v0.7.0)

Bidirectional Communication

The Control Protocol enables bidirectional communication between agents and clients for interactive workflows.

from kaizen.core.autonomy.control import ControlProtocol
from kaizen.core.autonomy.control.transports import CLITransport

# Create protocol with CLI transport
protocol = ControlProtocol(CLITransport())

# Use with agent
agent = BaseAgent(
    config=config,
    signature=signature,
    tools="all",  # Enable tools via MCP
    control_protocol=protocol  # Enable bidirectional communication
)

# Start protocol
import anyio
async with anyio.create_task_group() as tg:
    await protocol.start(tg)

    # Agent can now ask questions during execution
    answer = await agent.ask_user_question(
        "Which environment?",
        ["dev", "staging", "production"]
    )

    # Request approval for dangerous operations
    approved = await agent.request_approval(
        "Delete old files?",
        {"files": ["old1.txt", "old2.txt"], "count": 2}
    )

    # Report progress
    await agent.report_progress("Processing files", percentage=50)

Available Transports

• CLITransport: Interactive command-line interface
• HTTPTransport (SSE): Server-sent events for web UIs
• StdioTransport: Standard I/O for MCP integration
• MemoryTransport: In-memory for testing

🔧 Creating Custom Agents

Basic Custom Agent

from kaizen.core.base_agent import BaseAgent
from kaizen.signatures import Signature, InputField, OutputField
from dataclasses import dataclass

# 1. Define your configuration
@dataclass
class SentimentConfig:
    llm_provider: str = "openai"
    model: str = "gpt-4"
    temperature: float = 0.2
    categories: list = None  # Custom field

    def __post_init__(self):
        if self.categories is None:
            self.categories = ["positive", "negative", "neutral"]

# 2. Define your signature
class SentimentSignature(Signature):
    text: str = InputField(desc="Text to analyze")

    sentiment: str = OutputField(desc="Sentiment category")
    confidence: float = OutputField(desc="Confidence 0.0-1.0")
    explanation: str = OutputField(desc="Brief explanation")

# 3. Extend BaseAgent
class SentimentAgent(BaseAgent):
    def __init__(self, config: SentimentConfig):
        super().__init__(config=config, signature=SentimentSignature())
        self.sentiment_config = config

    def analyze(self, text: str) -> dict:
        """Analyze sentiment with domain-specific logic."""
        # BaseAgent.run() handles everything else
        result = self.run(text=text)

        # Add custom validation
        if result["sentiment"] not in self.sentiment_config.categories:
            result["warning"] = f"Unexpected category: {result['sentiment']}"

        return result

# Usage
config = SentimentConfig(llm_provider="openai", model="gpt-4")
agent = SentimentAgent(config)
result = agent.analyze("This product is amazing!")
print(result["sentiment"])  # "positive"
print(result["confidence"])  # 0.92

Advanced: Custom Strategy

from kaizen.strategies.base import Strategy
from typing import Dict, Any

class CustomStrategy(Strategy):
    """Custom execution strategy for specialized workflows."""

    async def execute(self, signature, inputs: Dict[str, Any], config) -> Dict[str, Any]:
        # Custom pre-processing
        processed_inputs = self.preprocess(inputs)

        # Execute with LLM
        result = await self.llm_call(signature, processed_inputs, config)

        # Custom post-processing
        return self.postprocess(result)

# Use in custom agent
class AdvancedAgent(BaseAgent):
    def __init__(self, config):
        super().__init__(
            config=config,
            signature=MySignature(),
            strategy=CustomStrategy()  # Use custom strategy
        )

🔌 Integration Patterns

Integration with DataFlow

from dataflow import DataFlow
from kaizen.agents import SimpleQAAgent
from kaizen.agents.specialized.simple_qa import SimpleQAConfig
from kailash.workflow.builder import WorkflowBuilder
from kailash.runtime.local import LocalRuntime

# DataFlow for database operations
db = DataFlow()

@db.model
class QASession:
    question: str
    answer: str
    confidence: float
    timestamp: str

# Kaizen for AI processing
agent = SimpleQAAgent(SimpleQAConfig(llm_provider="openai", model="gpt-4"))
result = agent.ask("What is the capital of France?")

# Store in database via DataFlow nodes
workflow = WorkflowBuilder()
workflow.add_node("QASessionCreateNode", "store", {
    "question": "What is the capital of France?",
    "answer": result["answer"],
    "confidence": result["confidence"],
    "timestamp": "2025-01-17T10:30:00"
})

runtime = LocalRuntime()
results, run_id = runtime.execute(workflow.build())

Integration with Nexus

from nexus import Nexus
from kaizen.agents import SimpleQAAgent
from kaizen.agents.specialized.simple_qa import SimpleQAConfig

# Create Nexus platform
nexus = Nexus(
    title="AI Q&A Platform",
    enable_api=True,
    enable_cli=True,
    enable_mcp=True
)

# Deploy Kaizen agent via Nexus
agent = SimpleQAAgent(SimpleQAConfig())
agent_workflow = agent.to_workflow()
nexus.register("qa_agent", agent_workflow.build())

# Agent now available on all channels:
# - REST API: POST /workflows/qa_agent
# - CLI: nexus run qa_agent --question "What is AI?"
# - MCP: qa_agent tool for AI assistants like Claude

🧪 Testing

3-Tier Testing Strategy

Kaizen uses a rigorous 3-tier testing approach with NO MOCKING in Tiers 2-3:

1. Tier 1 (Unit): Fast, mocked LLM providers (~450+ tests)
2. Tier 2 (Integration): Real Ollama inference (local, free)
3. Tier 3 (E2E): Real OpenAI inference (paid API, budget-controlled)

# Run all tests
pytest

# Run Tier 1 only (fast, mocked)
pytest tests/unit/

# Run Tier 2 (Ollama integration - requires Ollama running)
pytest tests/integration/test_ollama_validation.py

# Run Tier 3 (OpenAI - requires API key in .env)
pytest tests/integration/test_multi_modal_integration.py

Testing Custom Agents

import pytest
from kaizen.agents import SimpleQAAgent
from kaizen.agents.specialized.simple_qa import SimpleQAConfig

def test_simple_qa_agent():
    """Test basic Q&A functionality."""
    config = SimpleQAConfig(llm_provider="mock")  # Use mock provider for unit tests
    agent = SimpleQAAgent(config)

    result = agent.ask("What is 2+2?")

    assert "answer" in result
    assert "confidence" in result
    assert isinstance(result["confidence"], float)
    assert 0 <= result["confidence"] <= 1

def test_memory_enabled_agent():
    """Test memory continuity across sessions."""
    config = SimpleQAConfig(max_turns=10)  # Enable memory
    agent = SimpleQAAgent(config)

    # First interaction
    result1 = agent.ask("My name is Alice", session_id="test123")

    # Memory recall
    result2 = agent.ask("What's my name?", session_id="test123")

    assert "alice" in result2["answer"].lower()

📦 Production Deployment

Environment Configuration

# Required API Keys (.env file)
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...

# Optional Configuration
KAIZEN_LOG_LEVEL=INFO
KAIZEN_PERFORMANCE_TRACKING=true
KAIZEN_ERROR_HANDLING=true

Docker Deployment

FROM python:3.11-slim

# Install Kaizen
RUN pip install kailash-kaizen

# Copy application
COPY app.py .
COPY .env .

# Run application
CMD ["python", "app.py"]

Production Agent Configuration

from kaizen.agents import SimpleQAAgent
from kaizen.agents.specialized.simple_qa import SimpleQAConfig

# Production configuration with enterprise features
config = SimpleQAConfig(
    llm_provider="openai",
    model="gpt-4",
    temperature=0.1,  # Lower for consistency
    max_tokens=500,
    timeout=30,  # Request timeout
    retry_attempts=3,  # Retry on failures
    max_turns=50,  # Enable memory with limit
    min_confidence_threshold=0.7  # Quality gate
)

agent = SimpleQAAgent(config)

📊 Performance

BaseAgent Performance Improvements

• Async Execution: 2-3x faster than sync execution
• Lazy Loading: <100ms framework initialization
• Code Reduction: 87% less code vs traditional agents
• Auto-Optimization: Strategy-based execution optimization

Multi-Modal Performance

• Vision (Ollama): ~2-5 seconds per image (local, free)
• Vision (OpenAI): ~1-2 seconds per image (paid, higher quality)
• Audio (Whisper): ~0.5x real-time (1 min audio → ~30 sec processing)

📚 Examples

Complete Examples Repository

Kaizen includes 35+ working examples across 8 categories:

examples/
├── 1-single-agent/        # 10 basic patterns
│   ├── simple-qa/
│   ├── chain-of-thought/
│   ├── react-agent/
│   └── ...
├── 2-multi-agent/         # 6 coordination patterns
│   ├── supervisor-worker/
│   ├── consensus-building/
│   └── ...
├── 3-enterprise-workflows/ # 5 production patterns
│   ├── customer-service/
│   ├── document-analysis/
│   └── ...
├── 4-advanced-rag/        # 5 RAG techniques
│   ├── agentic-rag/
│   ├── graph-rag/
│   └── ...
├── 5-mcp-integration/     # 5 MCP patterns
├── 8-multi-modal/         # Vision/audio examples
└── README.md              # Examples overview

Location: /Users/esperie/repos/projects/kailash_python_sdk/apps/kailash-kaizen/examples/

⚠️ Common Mistakes

1. Missing .env Configuration

# ❌ WRONG: Not loading environment variables
from kaizen.agents import SimpleQAAgent
from kaizen.agents.specialized.simple_qa import SimpleQAConfig
agent = SimpleQAAgent(SimpleQAConfig(llm_provider="openai"))  # Fails!

# ✅ CORRECT: Load .env first
from dotenv import load_dotenv
load_dotenv()
agent = SimpleQAAgent(SimpleQAConfig(llm_provider="openai"))  # Works!

2. Wrong Vision Agent API

# ❌ WRONG: Using 'prompt' and 'response'
result = vision_agent.analyze(image=img, prompt="What is this?")
answer = result['response']

# ✅ CORRECT: Use 'question' and 'answer'
result = vision_agent.analyze(image="/path/to/image.png", question="What is this?")
answer = result['answer']

3. Using BaseAgentConfig Directly

# ❌ WRONG: Using BaseAgentConfig directly
from kaizen.core.config import BaseAgentConfig
config = BaseAgentConfig(model="gpt-4")  # Don't do this!

# ✅ CORRECT: Use domain config (auto-converted)
from kaizen.agents.specialized.simple_qa import SimpleQAConfig
config = SimpleQAConfig(model="gpt-4")
agent = SimpleQAAgent(config)  # Auto-extraction happens here

🔗 Additional Resources

Documentation

• CLAUDE.md - Quick reference for Claude Code
• Examples - 35+ working implementations
• Core SDK - Foundation patterns
• DataFlow - Database framework integration
• Nexus - Multi-channel platform integration

Guides

• Installation Guide - Setup and dependencies
• Quickstart Tutorial - Your first agent
• Signature Programming - Type-safe I/O
• BaseAgent Architecture - Unified system
• Multi-Modal Processing - Vision and audio
• Multi-Agent Coordination - A2A protocol

Reference

• API Reference - Complete API docs
• Configuration Guide - All config options
• Troubleshooting - Common issues

Community

• GitHub Repository - Source code and issues
• Kailash SDK Documentation - Main SDK documentation

---

Ready to get started? Begin with our Quickstart Tutorial or explore Working Examples.