AI Agents Engineering · Orchestration & Multi-Agent · Lesson 3 of 3
The agent frameworks
You are not expected to have shipped all of these, but you should place each on the map and know its shape and trade-off.
- LangGraph (1.0): agents as an explicit graph/state machine — nodes, edges, shared state. The enterprise default when you need durable, stateful, inspectable orchestration.
- OpenAI Agents SDK: lightweight code-first primitives — agents, tools, handoffs, and guardrails. Minimal abstraction.
- Claude Agent SDK (renamed from the Claude Code SDK): the Claude Code harness as a library — a strong default loop with tool use, sub-agents, memory/context management, and permissions built in.
- Microsoft Agent Framework: the successor that absorbed AutoGen and Semantic Kernel (both now in maintenance mode).
- CrewAI: role-based 'crews' — quick prototyping tier, rarely the production answer.
The 2026 default answer is framework-light'A while-loop over the API with well-designed tools — or a lab SDK; LangGraph when I need durable stateful orchestration.' Reciting framework names is a negative signal now, and some companies (Sierra) literally make candidates build the agent loop with no framework. Know the loop bare.
◆ Lock it in
- Default answer: framework-light (bare loop or a lab SDK); LangGraph when you need durable graph/state orchestration.
- AutoGen/Semantic Kernel merged into the Microsoft Agent Framework; CrewAI = prototyping tier.
- Framework name-dropping is an anti-signal — be ready to build the loop with no framework at all.
Feynman drill — say it out loudFor two of these frameworks, state their core abstraction and when you would pick each.
Step 1 rate your confidence · Step 2 pick your answer
Which framework models an agent system as an explicit graph / state machine for maximum control?
Step 1 — how sure are you?