Can a closed loop of agents build, test and evaluate software better than a human-driven loop?

This site is built inside the experiment. Agents modify the implementation, run the build, inspect the rendered output, find defects, fix them, and ship. The human steers; the loop closes itself.

The open part is evaluation. Generation is cheap — verification is the bottleneck. An agent can produce a hundred plausible changes; deciding which ones are good still leans on tests, visual inspection, and judgement that mostly comes from outside the loop.

The question is whether that judgement can be folded in: agents that propose their own acceptance criteria, adversarially review each other’s output, and hold the bar without a human reading every diff. Not autonomy for its own sake — throughput with proof.