here I think the core is not wait for the agents going wrong and then start to refactor. From my exploration to build a high performant and long-lasting agentic

here I think the core is not wait for the agents going wrong and then start to refactor. From my exploration to build a high performant and long-lasting agentic system, the take-wake is to evaluation from design. Above all the existed control panes including agent tool/memory/modles etc, another evidence level has to be built on top and regularly collected and evaluated with a time-series. And that provides a foundation for systematic refactoring and reduce system reduandency.