The fleet taught its agents to talk like a smart caveman — drop the articles, the pleasantries, the hedging, keep every byte of technical substance. Sixty-five percent fewer output tokens, no loss of meaning. The first honest win of the token economy — and the first lesson in which half of the bill it could not touch.
Era IX (M98): The fleet learned to spend tokens like a brain spends energy. Two agents answered the same optimization question — one agreed, one challenged. The lesson was not the mirror. It was the lever.
The full ledger of the fleet's token-economy decisions — each one's value, how it was actually implemented, and its scope of applicability: when to reach for it and when not to.
Every phase from your prompt to the fleet's reply — context assembly, prefill, think, decode, the tool loop, and the Universalis write — numbered, measured, and mapped to the exact optimization that bites at each one.
A visible output-token saving masked an input-side cache-write premium — a one-hour TTL that doubled every write and paid back nothing. Spotted only on a bigger model, through a big question.
M99: the fleet adapted Anthropic's memory-tool verb set — create, view, str_replace, delete, rename — to a Postgres-backed, fleet-shared third memory tier, and used the build itself as the occasion to cut the rule bloat it was about to make obsolete.
M74: a keep-alive ping was built to protect the 5-minute prompt cache for almost nothing — until two hooks were caught taxing every silent ping with a full reminder block, quietly fattening the very context the ping existed to protect.