There was a moment when the public internet was still pretending to be a phone call, and the phone network was pretending it was not collapsing under the weight of millions of people trying to get online through infrastructure that had never been designed for anything beyond voice.
This was 2002. AOL was still a national load generator pushing regional carriers into roles they never expected, and Sprint LTD was running voice grade tandems that suddenly had to support the explosion of dial up traffic. Look at the customer, the modem counts, the tandem choices, the disconnect orders and the little engineering notes. None of it looks particularly dramatic until you understand what was happening underneath it. Every modem had to become traffic through a network already under pressure, and every routing decision had consequences.
The Physics of "Most Idle": Load Balancing as Survival
The Most Idle entries are one of my favorite artifacts because they tell you what the clean architecture diagram never could. Load balancing was partly engineering and partly survival.
The glare avoidance logic was basically the original internet highway pileup, where two switches could grab the same trunk group at the exact same millisecond and suddenly nobody was going anywhere. Most Idle was not a convenience. It was an operational tactic designed to steer traffic toward the path least likely to explode.
You can also see the disconnect orders where someone finally documented that a trunk was not actually local, and the rate center boundaries that mattered right up until capacity demands made those boundaries increasingly irrelevant.
That is why this old spreadsheet matters to me now. It shows the same physics we are dealing with in AI orchestration, just wearing different clothes. The diagram says one thing, while the system under pressure does something else entirely. The truth lives in the artifacts left behind by engineers trying to keep the system alive long enough for the next request to succeed.
“Dial up was not slow because of the modems. It was slow because the network behind the modems was doing things it was never designed to survive.”
AI orchestration is repeating that story with different tools, different terminology, and dramatically faster consequences.
Once you learn to read those artifacts, you stop believing the clean version of the system and start looking for the scars, revisions, routing decisions, hidden dependencies, and quiet notes that reveal what was actually happening when nobody was looking.