workflow-skill

Create a node function that accepts and returns the workflow state TypedDict, then register it with graph.add_node and connect it using graph.add_edge or graph.add_conditional_edges. The node receives the full state and returns an updated state dictionary.

Model the plan as a list of PlanStep objects where each step declares its dependencies by node_id. The execute node runs steps whose dependencies are complete, executing independent steps in parallel and passing dependency results as context.

Yes, implement an evaluate node that checks goal achievement and routes back to the plan node via conditional edges when the goal is not met. Track iteration_count and max_iterations in state to prevent infinite replan loops.

Normal edges always route to a fixed next node, while conditional edges call a routing function that inspects the state and returns the next node's name. Use conditional edges after classify and evaluate nodes for intent-based or outcome-based branching.

Use a supervisor graph when tasks require multiple specialized agents, dependency ordering, parallel execution, or iterative refinement. Single agents suffice for casual chat or simple one-step requests that need no planning.