04
Frameworks & orchestration
Integration layer vs. orchestration engine
Verified LangChain is the integration layer; LangGraph is the durable orchestration engine (GA Oct 2025, in production at Uber, LinkedIn, Klarna, JPMorgan). Every major lab now ships its own agent SDK.
| Framework | Role & notes | Watch out for |
|---|---|---|
| LangGraph | Production default for complex workflows. Graph state, durable checkpointing, HITL interrupts. | Upfront state-modeling effort. |
| CrewAI | Fastest prototyping — working multi-agent in <20 lines. Abstractions: agents, tasks, crews, flows. | No built-in checkpointing; no authz model. Teams often re-implement in LangGraph for prod. |
| AutoGen / AG2 | Conversation-oriented multi-agent coordination. | AutoGen in maintenance mode (mid-2026); fragmented across AutoGen/AG2/MS forks. |
| OpenAI Agents SDK | Mar 2025, replaced Swarm. Handoffs, auto tool-schema gen. | Locked to OpenAI models. |
| Google ADK | Apr 2025. Native A2A for cross-framework agent comms. | — |
| Claude Agent SDK | Fastest-rising demand (~500× YoY search). Native MCP, Computer Use, hierarchical subagents (3 levels). | — |
| MS Agent Framework | Apr 2026 — merges Semantic Kernel + AutoGen. Native MCP + A2A. | Young / consolidating. |
Debunked
"Search volume proves LangGraph is most-adopted."
The 27,100-searches claim was refuted 0–3. "Dominant" reflects enterprise/citation signals — AutoGen leads raw GitHub stars.