Why per agent cost attribution is the dashboard panel that catches most multi-agent waste

When a multi-agent feature gets expensive, the first instinct is to look at the total bill. That number is alarming but unhelpful. It tells you something is wrong, not what to do. The single most useful panel is per-agent cost, shown as a stacked bar over time. Each agent in the team is a layer in the stack; each bar is a day or a deployment slice. As soon as you look at it, one layer will dominate. In almost every multi-agent system the dominant layer is the supervisor. The supervisor reads the whole conversation every turn to decide who speaks next, so its context grows linearly with the chat and its token bill grows with it. Sometimes the dominant layer is a critic that re-reads many draft revisions. Sometimes it is a specific worker that should have finished in one hop and quietly turned into three. The diagnostic flips the conversation. Instead of 'multi-agent is expensive', you can say 'the supervisor accounts for 47 percent of the bill; let us move it to a cheaper router model and summarise its input.' That sentence has a name and a fix. The bonus enrichment is slicing the per-agent panel by task type. Often you find that simple tasks burn the same multi-agent cost as hard tasks, which leads to a head-tail split: route the common easy tasks to a single agent and reserve the team for the long tail.

**Per-agent cost as a stacked bar over time** is the panel to ship first. Aggregate cost is a number; per-agent cost is a diagnosis. The value comes from two places. First, multi-agent cost distribution is almost never uniform. One agent dominates, usually the supervisor (because its context grows with the chat log every turn) or a critic loop (because it re-reads draft revisions). Aggregate cost hides that asymmetry; per-agent cost surfaces it on first glance. Second, each agent has a different cheapest fix. A supervisor that dominates because it carries the full chat is fixed by summarising its input; one that dominates because it is on a frontier model is fixed by demoting it to a router-grade model. A critic loop that dominates is fixed by capping max iterations or by making the score threshold easier to clear. None of these moves is visible from the aggregate view. The enrichment that often pays for itself is slicing the per-agent panel by task type. The common 'multi-agent is needed for the long tail and burning money on the head' pattern shows up clearly: simple queries cost the same as hard ones because they all go through the full team. The fix is a head-tail split, where a fast classifier sends easy tasks to a single agent and reserves the multi-agent team for the long tail. Both the diagnosis and the fix come from one panel.