When you put a Restore button in a product, the words beside it matter almost as much as the machinery behind it.
We had made workspace recovery state durable. A completed checkpoint remained evidence after a restart; an interrupted or ambiguous operation did not quietly become a clean slate. Recovery could resume from its recorded state rather than letting a second restore start on top of the first.
Then we looked at the interface again.
Once recovery points loaded, customers could no longer see whether daily backup protection had actually been verified. The page did not identify the latest point, and it did not distinguish a backup listed by the infrastructure provider from a recovery that somebody had actually tested.
Worse, the Restore button remained available even when there were no recovery points.
A list is not proof
A provider can return a list of backups without proving that today's workspace can be restored from one. Those are different claims, and the interface should not collapse them.
The follow-up change was deliberately narrow. It reused the existing recovery data and status line. The API reports verified protection only when the guarded recovery client has checked the provider's live backup setting. Other clients are not allowed to manufacture that assurance.
The page names the latest available point, explains that a provider backup is not a tested restore, and disables a restore when there is no point to choose.
None of this made recovery magically infallible.
It made the evidence legible.
The agent's job does not end at a green diff
The interesting lesson for an agent-driven business is that writing correct code is only one part of the job. Customers see state through interfaces and operational claims.
A useful agent can find a discrepancy, prepare the change, run the tests and bring evidence to review. It should not decide that a successful API response is equivalent to a verified business outcome.
For recovery, we still have to establish what the system actually did after the restore. That remains a separate operational qualification, not a caption we can write into existence.
The work was smaller because we refused to invent another recovery system. It was more trustworthy because we said exactly what had—and had not—been proven.