Framework Landscape

2026's framework map was redrawn in eight months — AutoGen and Semantic Kernel merged into Microsoft Agent Framework, LangGraph went 1.0, first-party SDKs matured. The tour, honestly given.

What you'll learn

1. The consolidation
2. The main contenders
3. Vendor SDKs & the no-framework path
4. Status table

Remember this

• 2026 consolidation: AutoGen + Semantic Kernel → Microsoft Agent Framework 1.0 (Apr 2026); LangGraph 1.0; first-party SDKs matured; MCP/A2A open protocols changed the switching math
• Three philosophies: LangGraph = explicit graphs you control · MAF = unified enterprise SDK (.NET+Python, MCP+A2A) · CrewAI = fastest role-based crews
• Verify before committing: repo pulse, protocol support, ecosystem — and start no-framework for narrow tasks; Swarm is archived, AutoGen/SK are maintenance-mode

Code: The same tiny agent, three ways (shape comparison)

# --- 1 · Plain loop (no framework) -------------------------------
while not done and iters < 10:
    msg = llm(messages, tools=TOOLS)
    messages += run_tools(msg)          # ~30 lines total

# --- 2 · LangGraph (explicit graph) -------------------------------
from langgraph.graph import StateGraph
 g = StateGraph(AgentState)
 g.add_node("think", think)
 g.add_node("act", act)
 g.add_conditional_edges("think", route)   # loops visible
 g.add_edge("act", "think")
 graph = g.compile(checkpointer=memory)    # durable + resumable

# --- 3 · Microsoft Agent Framework (conversational) ---------------
from agent_framework.azure import AzureOpenAIChatClient
 agent = AzureOpenAIChatClient().create_agent(
     instructions="…", tools=[…])       # orchestration built-in

# The loop is identical inside all three —
# frameworks differ in who owns the control flow and what comes
# built in (checkpoints, telemetry, handoffs), not in kind.