Checkpoint the step, not the pipeline: four agents and one unguarded embed

The first line of the workflow definition is the one doing the work. `initiateStateHydration` runs before any agent, and every stage after it is wrapped in a branch that fires only if a flag says the work is not already finished ^[8]. The pipeline does not remember what it did; it asks Postgres and gets booleans back ^[8][9]. That is checkpointed batch processing, and it predates the agents by decades. What the models change is the price of a lost step.

The second useful decision is that each sub-workflow is two steps rather than one: the agent call, then the write ^[9]. The agent writes into workflow state and the persistence step reads from it, so a failed insert can be retried without paying for the completion again ^[10]. Be precise about the scope of that. The post describes workflow state as an in-memory, transient object ^[10], so the failure it covers is a database write that fails inside a process still running. The failure the post opens with, a worker that crashes mid-pipeline ^[3], takes that state with it, and the step in flight is paid for twice ^[13].

Then the gap. Three of the four processing stages sit behind skip guards: parser, summary, risk ^[8]. The embedding workflow is chained with a plain `.then()` and carries no flag ^[12]. Every resumed run re-executes it. The post's own list of things that go wrong includes burning API quota re-processing steps that already succeeded ^[3], and embedding is that step, unguarded ^[12]. Whatever it costs, in tokens or in wall-clock across the chunks of a long lease, the resume path pays it in full each time.

It is also worth naming what a skip flag records. Zod validation proves shape, and the post is explicit that a step either validates or fails ^[6]. A parser that returns null for an expiry date it could not find is behaving correctly and validates ^[5]. So does a confidently wrong governing law. Once the stage is marked complete, the checkpoint that saves quota also shields that output from ever being recomputed, which means reprocessing after a prompt or model change needs a way to clear flags per stage. The text available to us stops before it gets there ^[11].

The same truncation covers the fourth failure mode. Contention between concurrent uploads for rate-limited endpoints is on the list of problems the design claims to have handled from the start ^[3], but no concurrency limit appears in the portion of the post we can read ^[15].

The granularity is the part to carry away. The smallest recoverable unit here is one sub-workflow, meaning one agent's entire call ^[14]. That sets a rule with nothing to do with AI: if a step is expensive enough that repeating it hurts, it has to be small enough that repeating it is cheap.