Framework / Nodes
Give every intelligent step one job
Choose a pure Python node when your code already knows the answer, and a model-backed node only for the narrow uncertainty that remains.Node types
Code
@node
Deterministic Python capability and the right home for application side effects.
Structure
SchemaNode
Extract or generate JSON that matches a declared output model.
Defined steps
DeterministicReasoningNode
Run an explicit ordered list of ReasoningStep responsibilities.
Bounded reasoning
StochasticReasoningNode
Let the model reason within one bounded node and its tool allowlist.
Reason + shape
CombineNode
Gather evidence with reasoning, then finish with a SchemaStep-constrained result.
Selection rule
Use deterministic code for policy, arithmetic, authorization, and side effects. Use SchemaNode for typed extraction. Use DeterministicReasoningNode when responsibilities are known in advance, StochasticReasoningNode for one bounded exploratory responsibility, and CombineNode when evidence gathering must end in a typed object.