01

Define completion

Choose one bounded workflow and define its task identity, states, outputs, owner, and observable completion criteria. Parallel activity is not the goal; a new worker must be able to continue after an interruption.

02

Draw dependencies

Draw dependencies before assigning roles. Research streams may run in parallel, while writing waits for facts and publishing waits for reviewed copy, images, and localization. Avoid having two agents mutate the same artifact at once.

03

Persist checkpoints

Save checkpoints outside the model conversation. Store completed actions, source and artifact URIs, model and prompt versions, open questions, the next step, and timestamps. Recovery begins by validating this state against real files and remote systems.

04

Make effects idempotent

Use stable idempotency keys for uploads, publishing, messages, database writes, and paid actions. A client timeout does not prove the server failed. Prepare, validate, and commit in separate phases, then store remote IDs after success.

05

Test recovery

Put human approval before irreversible or external actions. Test by killing a run, timing out after upload, and racing duplicate slugs. A second agent should recover without reading the full previous conversation.