“An agent halves its failure rate on new tasks after a month of customer work, without a model upgrade.”

Benaich frames this as a practical test of continual learning: evidence that customer work itself produces reusable learning rather than model upgrades.