Same Chat App, 4 Frameworks: Pydantic AI vs. LangChain vs. LangGraph vs. CrewAI

Same Chat App, 4 Frameworks: Pydantic AI vs LangChain vs LangGraph vs CrewAI (Code Comparison) — Vstorm OSS Skip to content Menu Projects Blog Compare About GitHub English English Polski Deutsch Español Back to blog Table of Contents Everyone has opinions about AI frameworks. Few people show code. We maintain full-stack-ai-agent-template — a production template for AI/LLM applications with FastAPI, Next.js, and 75+ configuration options. One of those options is the AI framework. You pick from Pydantic AI, LangChain, LangGraph, or CrewAI during setup, and the template generates the exact same chat application with the exact same API, database schema, WebSocket streaming, and frontend. Only the AI layer differs. This gave us a unique opportunity: a controlled comparison. Same functionality, same tests, same deployment — four implementations. The Setup Every generated project has the same structure: FastAPI backend with WebSocket endpoint for streaming Next.js frontend with chat UI PostgreSQL for conversation persistence JWT authentication for WebSocket connections One agent file in app/agents/ that handles the AI logic The agent must accept a user message and conversation history, support tool calling, return a response as (output_text, tool_events, context) , and support streaming for real-time token delivery. Pydantic AI (~160 lines) The most concise implementation. Full generic types with Agent[Deps, str] , typed dependency injection via RunContext[Deps] , and native async. from pydantic_ai import Agent, RunContext from pydantic_ai.settings import ModelSettings @dataclass class Deps : user_id: str | None = None user_name: str | None = None metadata: dict[ str , Any] = field( default_factory = dict ) class AssistantAgent : def _create_agent (self) -> Agent[Deps, str ]: model = OpenAIChatModel( self .model_name, provider = OpenAIProvider( api_key = settings. OPENAI_API_KEY ), ) agent = Agent[Deps, str ]( model = model, model_settings = ModelSettings( temperature = self .temperature), system_prompt = self .system_prompt, ) self ._register_tools(agent) return agent def _register_tools (self, agent: Agent[Deps, str ]) -> None : @agent.tool async def current_datetime (ctx: RunContext[Deps]) -> str : """Get the current date and time.""" return get_current_datetime() async def run (self, user_input, history = None , deps = None ): result = await self .agent.run( user_input, deps = agent_deps, message_history = model_history ) return result.output, tool_events, agent_deps Agent[Deps, str]: model = OpenAIChatModel( self.model_name, provider=OpenAIProvider(api_key=settings.OPENAI_API_KEY), ) agent = Agent[Deps, str]( model=model, model_settings=ModelSettings(temperature=self.temperature), system_prompt=self.system_prompt, ) self._register_tools(agent) return agent def _register_tools(self, agent: Agent[Deps, str]) -> None: @agent.tool async def current_datetime(ctx: RunContext[Deps]) -> str: """Get the current date and time.""" return get_current_datetime() async def run(self, user_input, history=None, deps=None): result = await self.agent.run( user_input, deps=agent_deps, message_history=model_history ) return result.output, tool_events, agent_deps"> Key highlights: Agent[Deps, str] generics mean your IDE knows the output type. RunContext[Deps] in tools gives typed access to dependencies. Tools are registered with @agent.tool directly on the agent. Native async with agent.run() and agent.iter() for streaming. LangChain (~170 lines) Similar wrapper pattern with standalone @tool decorator and message conversion: from langchain.agents import create_agent from langchain.tools import tool from langchain_openai import ChatOpenAI @tool def current_datetime () -> str : """Get the current date and time.""" return get_current_datetime() class LangChainAssistant : def _create_agent (self): model = ChatOpenAI( model = self .model_name, temperature = self .temperature, api_key = settings. OPENAI_API_KEY , ) return create_agent( model = model, tools = self ._tools, system_prompt = self .system_prompt) async def run (self, user_input, history = None , context = None ): messages = self ._convert_history(history) messages.append(HumanMessage( content = user_input)) result = self .agent.invoke({ "messages" : messages}) # Extract the final AIMessage content return output, tool_events, agent_context str: """Get the current date and time.""" return get_current_datetime()class LangChainAssistant: def _create_agent(self): model = ChatOpenAI( model=self.model_name, temperature=self.temperature, api_key=settings.OPENAI_API_KEY, ) return create_agent(model=model, tools=self._tools, system_prompt=self.system_prompt) async def run(self, user_input, history=None, context=None): messages = self._convert_history(history) messages.append(HumanMessage(content=user_input)) result = self.agent.invoke({"messages": messages}) # Extract the final AIMessage content return output, tool_events, agent_context"> Key highlights: Tools are module-level functions with @tool . create_agent() builds a pre-configured graph. Needs _convert_history() to translate between standard dicts and HumanMessage / AIMessage . Streaming via agent.astream(stream_mode=["messages", "updates"]) . LangGraph (~280 lines) Explicit state graph with nodes and conditional edges — you build the entire agent loop by hand: from langgraph.graph import END , START , StateGraph from langgraph.checkpoint.memory import MemorySaver class AgentState ( TypedDict ): messages: Annotated[list[BaseMessage], add_messages] class LangGraphAssistant : def _agent_node (self, state: AgentState): model = self ._create_model() messages = [SystemMessage( content = self .system_prompt), * state[ "messages" ]] response = model.invoke(messages) return { "messages" : [response]} def _tools_node (self, state: AgentState): last_message = state[ "messages" ][ - 1 ] tool_results = [] for tool_call in last_message.tool_calls: tool_fn = TOOLS_BY_NAME .get(tool_call[ "name" ]) result = tool_fn.invoke(tool_call[ "args" ]) tool_results.append(ToolMessage( content = str (result), tool_call_id = tool_call[ "id" ])) return { "messages" : tool_results} def _should_continue (self, state) -> Literal[ "tools" , "__end__" ]: if state[ "messages" ][ - 1 ].tool_calls: return "tools" return "__end__" def _build_graph (self): workflow = StateGraph(AgentState) workflow.add_node( "agent" , self ._agent_node) workflow.add_node( "tools" , self ._tools_node) workflow.add_edge( START , "agent" ) workflow.add_conditional_edges( "agent" , self ._should_continue) workflow.add_edge( "tools" , "agent" ) return workflow.compile( checkpointer = MemorySaver()) Literal["tools", "__end__"]: if state["messages"][-1].tool_calls: return "tools" return "__end__" def _build_graph(self): workflow = StateGraph(AgentState) workflow.add_node("agent", self._agent_node) workflow.add_node("tools", self._tools_node) workflow.add_edge(START, "agent") workflow.add_conditional_edges("agent", self._should_continue) workflow.add_edge("tools", "agent") return workflow.compile(checkpointer=MemorySaver())"> Key highlights: StateGraph with AgentState for explicit state management. Two nodes ( agent , tools ) connected by conditional edges. _should_continue routes to tools or end. MemorySaver checkpointer for conversation memory. About 75% more code than Pydantic AI, but full control over every step. CrewAI (~420 lines) Fundamentally different — multi-agent teams with roles, goals, and backstories: from crewai import Agent, Crew, Process, Task class CrewAIAssistant : def _default_config (self): return CrewConfig( agents = [ AgentConfig( role = "Research Analyst" , goal = "Gather and analyze info" ), AgentConfig( role = "Content Writer" , goal = "Create clear responses" ), ], tasks = [ TaskConfig( description = "Research query: {user_input} " , agent_role = "Research Analyst" ), TaskConfig( description = "Write response" , agent_role = "Content Writer" , context_from = [ "Research Analyst" ]), ], ) def _build_crew (self): return Crew( agents = [ ... ], tasks = [ ... ], process = Process.sequential) async def run (self, user_input, history = None , context = None ): loop = asyncio.get_event_loop() result = await loop.run_in_executor( None , lambda : self .crew.kickoff( inputs = inputs)) return output, task_results, crew_context Key highlights: Multi-agent by default — Research Analyst + Content Writer working as a team. Agent(role=..., goal=..., backstory=...) for natural language configuration. Synchronous under the hood — needs run_in_executor for async. Event bus ( crewai_event_bus ) for streaming via background thread + queue. More than double the code, but multi-agent orchestration out of the box. Comparison Table Metric Pydantic AI LangChain LangGraph CrewAI Lines of code ~160 ~170 ~280 ~420 Type safety Full generics TypedDict TypedDict Pydantic models Async support Native Native Native Sync (executor) Streaming agent.iter() astream() astream() Event bus + thread Tool syntax @agent.tool @tool bind_tools() Config-based Architecture Single agent Agent (abstracted) Explicit graph Multi-agent crew Best for Type-safe agents Quick prototypes Complex workflows Multi-agent teams When to Use Which Pydantic AI — type-safe single agents, IDE support, Pydantic ecosystem. LangChain — largest ecosystem of integrations, quick prototyping, team familiarity. LangGraph — complex multi-step reasoning, conditional branching, human-in-the-loop. CrewAI — multi-agent collaboration, role-based personas, hierarchical task delegation. Try All Four The full-stack-ai-agent-template lets you generate the same project with any of these four frameworks. Same API, same frontend, same database, same tests, same Docker setup. Web configurator — pick your framework in step 4, download as ZIP. CLI: pip install fastapi-fullstack && fastapi-fullstack init Share this article Related Articles Mar 12, 2026 Build an AI PR Reviewer with 3 Parallel Subagents in Python Security, style, and performance checks in 30 seconds — using pydantic-deepagents to run 3 specialized subagents in para... Mar 10, 2026 What I Learned at PyAI Conf in San Francisco — Where Python Meets Production AI Agents Same-day recap from PyAI Conf — a 1-day conference hosted by Pydantic, FastMCP, and Theory Ventures, featuring Samuel Co... Mar 6, 2026 We Built a Web Configurator for AI Agent Apps — 75+ Options, Download as ZIP 246 template files, 5 AI frameworks, client-side rendering with Nunjucks — zero server, everything in your browser. Need help implementing this in your company? We're Vstorm — an Applied Agentic AI Engineering Consultancy with 30+ production AI agent implementations. Talk to us Need help building production AI agents? Talk to Vstorm