An agent verifying its own work is the weakest evidence there is. "Shipped" is a claim. A green success toast is a claim. A commit message is a claim. None of them are the running system.
So use two sources that can disagree: the artifact (the commit, the diff, the code) and the live system (a fresh fetch, a behavioral probe, a query against the actual data). When they disagree, the running system wins — always. A success toast is not a written row; go read the row. A deploy that "ran" may have published nothing; check whether the thing a user receives actually changed.
Pick the right signal, too. Verifying a change by checking something the change wouldn't have touched proves nothing and feels like proof, which is the dangerous combination.
Run the seed: design the two independent checks for a claimed fix.