What's the difference between a LangGraph workflow and a LangGraph agent?

A workflow’s control flow is code you wrote; an agent’s control flow is a decision the model makes at runtime. In a LangGraph workflow, edges are predetermined: you decided in advance which node runs after which, so the graph always follows one of a fixed set of paths. In an agent, the model itself decides which tool to call or which node to go to next, so the same graph can take a path nobody wrote down in advance.

That difference is what changes what needs grading. A workflow’s paths are enumerable, so you can write one eval per path and know you’ve covered the graph. An agent adds a decision you can’t enumerate in advance: whether the model picked the right next step given what it had seen so far, the same kind of check a mis-routed handoff needs in any multi-agent system. Grading a workflow means checking each node’s output; grading an agent means also checking the routing decision itself.