A small assistant deployment checklist

Start with one job and one owner. Write an example input and the exact result that would count as success. Choose a workload small enough that a person can review every outcome during the first trial.

Check timeouts, credentials, and the dependency failure path before adding more tasks. A job should stop clearly when required information is missing. It should not fill a missing account number or recipient with a plausible guess.

Capture action receipts and make reruns explicit. If the job creates a record, save its identifier. If it sends a request, save the intended action identifier and response state. These records help distinguish work that failed from work whose response was lost.

Increase volume after reviewing real runs. Measure completion, duplicate actions, unresolved exceptions, and the time a person spends fixing them. A deployment that saves ten minutes but needs an hour of recovery has not passed its trial.