Nonsense indeed. The model knowledge is the current state of the art. Any computation it does, advances it. It re-ingests work of prior agents every time you run it on your codebase, so even though the model initializes the same way (until they update the model), upon repeated calls it ingests more and more novel information, inching the state of the art ever forwards.