Frameworks Deep Dive — Multi-Agent Systems

CrewAI Flows vs. LangGraph v1.x vs. Microsoft Agent Framework vs. Claude Agent SDK

CrewAI (Crews + Flows) pairs role-based autonomous teams (Agent with role, goal, backstory, Task, and memory=True) with event-driven Flows using @start(), @listen(), and @router() decorators over typed Pydantic state:

from typing import Literal
from pydantic import BaseModel
from crewai.flow.flow import Flow, start, listen, router

class AuditState(BaseModel):
    risk_score: float = 0.0

class ComplianceFlow(Flow[AuditState]):
    @start()
    def ingest_report(self):
        self.state.risk_score = 0.82
        return self.state.risk_score

    @router(ingest_report)
    def route_by_risk(self, score: float) -> Literal["escalate", "approve"]:
        return "escalate" if score > 0.7 else "approve"

    @listen("escalate")
    def run_audit_crew(self):
        return audit_crew.kickoff()  # Triggers multi-agent Crew with memory

LangGraph models agents as state machines using directed graphs. State flows through nodes (agents/functions) connected by edges (conditional logic). It is harder to learn but the gold standard for production because it allows deterministic control flows and easy persistence (saving/resuming state).

Microsoft Agent Framework (microsoft/agent-framework, pip install agent-framework) is the official enterprise successor unifying AutoGen (v0.4+ async autogen-agentchat with AssistantAgent and RoundRobinGroupChat) and Semantic Kernel into one Python & .NET SDK with ChatAgent, @ai_function tools, graph workflows, native MCP/A2A support, and Entra ID governance:

from agent_framework import ChatAgent, ai_function
from agent_framework.openai import OpenAIChatClient

@ai_function
def lookup_sla(tier: str) -> str:
    """Return SLA response target for customer tier."""
    return "15m" if tier == "enterprise" else "4h"

agent = ChatAgent(
    client=OpenAIChatClient(model_id="gpt-6-astra"),
    name="SlaCoordinator",
    instructions="Verify customer SLA commitments accurately.",
    tools=[lookup_sla],
)

Anthropic Claude Agent SDK (claude-agent-sdk in Python and @anthropic-ai/claude-agent-sdk in TypeScript - renamed from claude-code-sdk / @anthropic-ai/claude-code) embeds Claude Code's autonomous tool-use and context-compaction loop directly via query(), ClaudeSDKClient, and ClaudeAgentOptions.