AI Agent Frameworks in 2025: LangGraph vs CrewAI vs AutoGen vs Raw API

// Raw Anthropic/OpenAI API: maximum control, no abstraction overhead
// Best for: simple agents, high-reliability production systems, cost optimization

// Complete agent in ~50 lines
async function codeReviewAgent(pr_diff) {
  const messages = [{
    role: 'user',
    content: 'Review this PR diff and identify: bugs, security issues, performance problems.\n' + pr_diff
  }];

  const resp = await anthropic.messages.create({
    model: 'claude-opus-4-5',
    max_tokens: 2000,
    tools: [reviewTool, searchTool],
    messages
  });
  // Handle tool calls, return results
  // 0 dependencies, fully debuggable, cheapest approach
}

// Pros: No abstraction leakage, full control, easy debugging
// Cons: More boilerplate, roll your own memory + orchestration
// Best for: teams that need to understand every decision the agent makes

// LangGraph: agent as state machine with explicit nodes + edges
// Best for: complex multi-step workflows, production APIs, resumable execution

import { StateGraph, END } from '@langchain/langgraph';

// Define state schema
const agentState = { messages: null, plan: null, step: 0, results: {} };

const graph = new StateGraph({ channels: agentState })
  .addNode('plan', planNode)          // Generate execution plan
  .addNode('execute', executeNode)    // Execute one step
  .addNode('verify', verifyNode)      // Check step output
  .addNode('error', errorNode)        // Handle failures
  .addEdge('plan', 'execute')
  .addConditionalEdges('verify', routeAfterVerify)  // Branch on state
  .setEntryPoint('plan');

// Key LangGraph advantages:
// - Checkpointing: save/restore agent state mid-execution
// - Time travel: replay from any checkpoint for debugging
// - Streaming: stream intermediate state to frontend
// - Human-in-the-loop: pause at any node for human approval
const app = graph.compile({
  checkpointer: new SqliteSaver(db)  // Persist state between runs
});

// CrewAI: define agents with roles, delegate tasks between them
// Best for: workflows that map naturally to human team structures

from crewai import Agent, Task, Crew, Process

researcher = Agent(
    role='Senior Research Analyst',
    goal='Find and synthesize the latest information on {topic}',
    backstory='Expert at finding authoritative sources and summarizing technical content',
    tools=[web_search_tool, arxiv_tool],
    llm='claude-opus-4-5',
    verbose=True
)

writer = Agent(
    role='Technical Content Writer',
    goal='Write clear, accurate technical content based on research',
    backstory='Experienced developer turned writer, expert at code examples',
    llm='claude-opus-4-5'
)

research_task = Task(
    description='Research the latest advances in {topic}. Find 5 key papers.',
    agent=researcher,
    expected_output='5 paper summaries with key findings'
)

writing_task = Task(
    description='Write a technical blog post based on the research',
    agent=writer,
    context=[research_task],  # Receives research output
    expected_output='1500 word blog post with code examples'
)

crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, writing_task],
    process=Process.sequential
)
result = crew.kickoff(inputs={'topic': 'AI agent memory architectures'})