I built an agent that improves its own pipeline, not just one that completes tasks

Most agent demos stop at "it did the task." I wanted to see what happens when you close the loop all the way, so I built a system that runs a live product end to end and then improves itself. It observes its own results, decides what to try next, acts by opening pull requests against both the product and its own code, and verifies each change through a ledger that marks it confirmed, rejected, or inconclusive. A human merge gate is the only manual step. The thing that made it work was resisting complexity. I kept context per run small, leaned on plain deterministic code for anything that didn't need judgment, and skipped the giant tool belt and long skill files entirely. The verify step turned out to be where the real value lives. Without forcing a confirmed, rejected, or inconclusive call on every change, the system just accumulates edits it can't reason about later, and the self-improving part quietly stops being true. For anyone building autonomous agents, how are you closing the verify step so the system actually learns instead of just acting?