TLDR; The core idea described in article is to identify components that makes things "hard" and then apply each component in isolation to "easy" version and master that easy+one hard thing problem. After you do that for each component, you likely find thing isn't hard anymore. This is "single loop system". The "double loop system" applies to thing that you don't know what the "good" final state looks like. So you basically consider that as problem and apply single loop system to first find out the good final state. Then again apply single loop system to achieve that good final state.
This is insightful but claims are rather over-inflated. More formally thinking, if this system works then my first instinct would be to apply on np-hard problems or build machine with human level intelligence. The issue is that hard things are often hard because solution search space is extremely vast and achieving any level of predictability is difficult because there is almost always new information lurking in space you haven't explored yet.
To be fair, the article does explicitly call out that the system won't work for arbitrarily hard or impossible problems. That's addressed in the opening paragraphs, and again called out in step 4.4 of the single-loop.
This is insightful but claims are rather over-inflated. More formally thinking, if this system works then my first instinct would be to apply on np-hard problems or build machine with human level intelligence. The issue is that hard things are often hard because solution search space is extremely vast and achieving any level of predictability is difficult because there is almost always new information lurking in space you haven't explored yet.