Framework Integration Overview

Agent Kernel supports multiple AI agent frameworks through a unified adapter pattern.

Supported Frameworks

Framework Comparison

Framework	Best For	Complexity	Multi-Agent
OpenAI Agents	Production apps with OpenAI	Low	Yes
CrewAI	Role-based collaboration	Medium	Yes
LangGraph	Complex workflows	High	Yes
Google ADK	Google ecosystem	Low	Yes
Smolagents	Lightweight tool-driven agents	Medium	Yes
Pydantic AI	Multi-provider apps, provider failover, typed structured output	Low	Yes
LiveKit Agents (coming soon)	Real-time voice/video agents	Medium	Yes

Capability Matrix

Not every Agent Kernel capability is available on every framework:

Capability	OpenAI Agents	CrewAI	LangGraph	Google ADK	Smolagents	Pydantic AI
Token streaming (execution.mode: stream)	✅	❌	✅	✅	❌	✅
Structured output (AgentReplyAny)	✅ output_type	✅ module-level output_pydantic/output_json	✅ response_format	✅ output_schema	✅ dict/Pydantic final_answer	✅ output_type
Portable tools (ToolBuilder)	✅	✅	✅	✅	✅	✅
Multimodal attachments	✅	✅*	✅*	✅	✅*	✅*
Hooks / guardrails / sessions / deployments	✅	✅	✅	✅	✅	✅
Native multi-provider models	⚠️ OpenAI-first	via LiteLLM	via LangChain	via LiteLLM	via LiteLLM	✅ many providers + FallbackModel

* Multimodal preprocessing (the attachment pre-hook and analyze_attachments tool) is framework-agnostic; direct image/file input via the REST API is currently documented for OpenAI Agents SDK and Google ADK. CrewAI and Smolagents raise NotImplementedError for streaming; use the default synchronous mode (or rest_sync on AWS) with those frameworks.

Choosing a Framework

OpenAI Agents SDK

• Official OpenAI support
• Simple API
• Built-in function calling
• Good for production

Learn more →

CrewAI

• Role-based agents
• Built-in collaboration patterns
• Easy task delegation
• Great for teams

Learn more →

LangGraph

• Graph-based orchestration
• Maximum flexibility
• Complex state management
• Best for sophisticated workflows

Learn more →

Google ADK

• Gemini models
• Google Cloud integration
• Simple agent creation
• Good for Google ecosystem

Learn more →

Smolagents

• Lightweight and composable design
• ToolCalling and CodeAgent support
• Managed-agent routing patterns
• Great for fast experimentation and mixed-agent setups

Learn more →

Pydantic AI

• Native multi-provider models (OpenAI, Anthropic, Google, Bedrock, Groq, Mistral, …)
• FallbackModel for automatic provider failover
• Typed, self-correcting structured output via output_type
• Delegation-via-tool for multi-agent routing (no handoffs= primitive)

Learn more →

LiveKit Agents (coming soon)

• Real-time audio and video agent framework
• Voice-enabled AI applications
• Low-latency media pipelines
• Ideal for conversational voice assistants and live-streaming AI bots

Migration Between Frameworks

Agent Kernel makes it easy to migrate:

# Original CrewAI implementation
from crewai import Agent
from agentkernel.crewai import CrewAIModule

agent = Agent(role="assistant", ...)
CrewAIModule([agent])

# Migrate to OpenAI (change 2 lines)
from agents import Agent
from agentkernel.openai import OpenAIModule

agent = Agent(name="assistant", ...)
OpenAIModule([agent])

Your execution code (CLI, API, deployment) remains unchanged!

Portable Tool Functions

Agent Kernel lets you write tool functions as plain Python and bind them to any framework using a ToolBuilder. The same tool works across all supported frameworks:

def get_weather(city: str) -> str:
    """Returns the weather for a given city."""
    return f"Weather in {city}: sunny"

# Same function, any framework
from agentkernel.openai import OpenAIToolBuilder
from agentkernel.crewai import CrewAIToolBuilder
from agentkernel.smolagents import SmolagentsToolBuilder

openai_tools = OpenAIToolBuilder.bind([get_weather])
crewai_tools = CrewAIToolBuilder.bind([get_weather])
smolagents_tools = SmolagentsToolBuilder.bind([get_weather])

Learn more about Tools →

Framework Portability

Available soon!

Features:

• Switch the underlying agentic framework without affecting the agent logic
• Effortless migration of already existing agents to the unified portable implementation