Cheat sheetAIS-02

Agent Frameworks

AI Specialization / Agent Frameworks

Frameworks give you retries, state, tracing, and multi-agent coordination for free. Pick by shape — graph, roles, conversation, or thin SDK — or use none.

LangGraphModel the agent as a graph of nodes/edges with persisted state. Best for loops, checkpoints, and durable human-in-the-loop.

CrewAIRole-based crews: define agents with goals ('researcher', 'writer') and tasks. Fast to prototype multi-agent collaboration.

AutoGenConversation-centric multi-agent: agents (and code executors) message each other in a loop until the task is done.

Supervisor/workerDominant multi-agent pattern: supervisor decomposes and routes sub-tasks to focused workers, then synthesizes results.

⚑

Start single-agent with the provider SDK. Only move to a framework when you hit a concrete need: durable state (LangGraph), quick multi-role setup (CrewAI), open agent dialogue (AutoGen), or provider-native tracing (Agents SDK).

Framework earns its keepA long-running research agent that must pause for human approval and resume after a restart — LangGraph's persisted state and checkpoints fit.

Framework is overkillA classify-then-route helper with one tool — a plain loop plus the provider SDK is clearer and easier to debug.

LangGraph: nodes + edges + shared persisted state
supervisor -> route sub-tasks -> workers -> synthesize
single-agent default; go multi only when sub-tasks separate

langgraphcrewaiautogenagents-sdkmulti-agentsupervisor-worker

review in 6d

Synced · Neon

AI Specialization / Agent Frameworks / AIS-02

Agent Frameworks

Frameworks give you retries, state, tracing, and multi-agent coordination for free. Pick by shape — graph, roles, conversation, or thin SDK — or use none.

Key points

LangGraphModel the agent as a graph of nodes/edges with persisted state. Best for loops, checkpoints, and durable human-in-the-loop.

CrewAIRole-based crews: define agents with goals ('researcher', 'writer') and tasks. Fast to prototype multi-agent collaboration.

AutoGenConversation-centric multi-agent: agents (and code executors) message each other in a loop until the task is done.

Supervisor/workerDominant multi-agent pattern: supervisor decomposes and routes sub-tasks to focused workers, then synthesizes results.

⚑

Examples

Framework earns its keepA long-running research agent that must pause for human approval and resume after a restart — LangGraph's persisted state and checkpoints fit.

Framework is overkillA classify-then-route helper with one tool — a plain loop plus the provider SDK is clearer and easier to debug.

Reference

LangGraph: nodes + edges + shared persisted state
supervisor -> route sub-tasks -> workers -> synthesize
single-agent default; go multi only when sub-tasks separate

langgraphcrewaiautogenagents-sdkmulti-agentsupervisor-worker

On this pageDefinition Key points Examples Reference

This topicnewReview in 6d

0% mastery · AI Specialization