You assume that the agents can produce okayish software that isn't very resilient, but that's not quite what we're seeing. Remember Anthropic's attempt to have agents write a C compiler (which isn't a super-complicated task by any means). Despite preparation work that is well beyond what most software can have (specs available to the agent and that the model trained on, thousands of human-written tests available to the agent and that the model trained on, and a human-written reference implementation available to the agent and that the model trained on) the agents still failed to converge. In other words, under conditions that are so favourable as to be unrealistic for almost any kind of real software, the AI agents couldn't even produce so much a workable C compiler. And this is what I also see: the agents take the code in a direction that doesn't converge, let alone make it resilient.