Agentic development allows one Claude, multiplexed a few times, to vastly improve its output and tackle much bigger problems than just prompting the one instance. If you had a million Claudes in layered networks like we do with matmuls to form Claude, you'd be really cooking with gas.
The output undoubtedly improves when looping LLM output back into the model at inference-time, but there is a limit to this and it is still bounded by the acumen of the underlying model. You cannot just recurse these models with tooling and compute to e.g. solve new physics.
(Maybe that's why they call it Gas Town?)