--- title: "What belongs on an agent-user dashboard" description: "Five task counts belong on the agent-user board. Traces, tokens, simulations, and human funnels do not share those tiles." canonical: "https://www.agentrail.co/blog/agent-user-analytics" author: "Vivek Mittal" published: "2026-11-12T12:11:00Z" last_updated: "2026-11-13T08:54:10.911Z" --- # What belongs on an agent-user dashboard Agent-user analytics on a dashboard is a handful of task counts: started, auth, first success, finished, and unfinished. Leave traces, tokens, and human funnels off that board. ![A white dashboard icon on a dark teal gradient](https://images.ctfassets.net/1htfyfnn1o5y/5nNdXvecOXu5TiZwJZfRm3/8f16a99b2fe2d79f3259459c345bf90b/22841f0c-635f-41c1-971b-2e202249bd85.png?w=1600&fm=webp&q=72) Agent-user analytics on a dashboard is a set of task counts. Five numbers earn a tile: tasks started, tasks that reached auth, tasks with a first successful call, tasks that finished, and tasks that did not finish. Readers will treat another chart on this board as one of those counts. The split between a log and a task stays on [agent observability vs agent experience](https://www.agentrail.co/blog/agent-observability-vs-agent-experience). The measure names stay on [how to measure agent experience](https://www.agentrail.co/blog/how-to-measure-agent-experience). > **Key Takeaways** > > The tiles are task counts: started, reached auth, first success, finished, and unfinished. > > Open a trace from a task. Keep the trace off the primary tiles. > > A retry is still one task. A simulation pass stays off the live tiles. > > Leave tokens, model charts, and human signup funnels off this board. ## Count tasks in one labeled window A tile on this dashboard is a count of tasks in a window you choose, such as seven days. A task is the unit. A span, a log line, and a model call are evidence attached to a task. When a count moves, open that task in [agent journey analytics](https://www.agentrail.co/blog/agent-journey-analytics), then open its trace. Request totals rise when one task retries, so the board can show more traffic while finished tasks stay flat. Count tasks on this board. Keep request volume in the trace tool, or on a panel labeled as requests. Use the same window on all five tiles, and label each tile with it. A month of finished tasks beside a day of first requests mixes two windows. Seven days is a readable default for a weekly review. Skip a sparkline of every HTTP status here. Status codes mix tasks that are in different stages. The five counts show the stage. ## Five counts earn a tile | Tile | What it counts | | --- | --- | | Started | Tasks that made a first request | | Reached auth | Tasks that presented a credential or hit a gate | | First success | Tasks that got a result they could use | | Finished | Tasks that met the outcome rule | | Unfinished | Tasks that started and did not finish | What counts as a request, as auth, and as a usable call stays on [the agent-user funnel](https://www.agentrail.co/blog/agent-user-funnel). The finished count uses [agent task completion rate](https://www.agentrail.co/blog/agent-task-completion-rate). Unfinished is that same outcome rule, as the tasks in the window that have no finished result. Put unfinished beside finished. A small finished count is hard to read until you see how many tasks started. A percentage can move because the denominator moved. The two counts show which side moved. Finished plus unfinished equals started, once every started task has an outcome. Auth and first success are subsets of started. If the tiles don't reconcile, one of them is counting a different unit. Fix that unit before you talk about a trend. Add an in-progress tile only when the outcome rule has a real open state that is not yet unfinished. When silence means the task is over, put it in unfinished once your timeout says so. That timeout lives with the outcome rule. Until you have one, a task in an unlabeled middle drops out of both finished and unfinished, and the tiles stop adding up. ## Leave traces, tokens, and human funnels off this board Keep these charts on another board. - Latency percentiles and error rate charts serve the on-call rotation. - Token totals and cost per model serve the person paying the model bill. - A leaderboard of models or agent products belongs on another board. - Citation share, mention counts, and search visibility are not task counts. - A human signup or checkout funnel follows a person. - Endpoint failure rankings are a route view. On this screen they crowd out the five tiles. Keep one filter. Restrict the five tiles to a single task name or a single API host. The filter narrows which tasks the same five counts include, so a docs-only task and a billing task do not share one drop. Put a chart here when the unit is an agent task that made a first request. A web-session chart renamed as agent sessions is still a human funnel. Traces stay one click from a task id. If the landing view is a list of spans, readers chase span volume. Land on the five counts, so the question is which stage lost tasks. ## A retry is still one task Increment the started tile on the first request, not on the twelfth. Counting each request as a task makes the auth tile and the finished tile impossible to compare. Treat the repeated calls as one stuck attempt on [the agent retry loop](https://www.agentrail.co/blog/agent-retry-loop). Link a task to that count if you need it, and leave the retry total off this board. A high retry total with a safe side effect is a different problem from a high unfinished count. A retry fix shouldn't show up as a change in finished tasks when the outcome didn't change. Five tool calls in one attempt are one task and five lines in the trace. This board assumes that grouping is already done. A simulation is a run you started. A live task showed up in traffic. They can share an outcome rule and still need separate tiles. A lab pass added to the finished count makes a quiet production week look healthy. A failed simulation on the unfinished tile changes that count when someone runs a suite. How those two charts compare stays on [agent simulation vs production analytics](https://www.agentrail.co/blog/agent-simulation-vs-production-analytics). Filter this board to live tasks, and put that in the title. ## Open the stage that lost the most tasks Once a week, read the five tiles for the same window. Find the neighboring gap that lost the most tasks: started to auth, auth to first success, or first success to finished. Unfinished should equal the tasks that fell in those gaps. If it doesn't, fix the accounting before you pick a project. If all five tiles are zero, check the collector before you sketch a funnel. Then open one unfinished task in the largest gap and read the call where it stalled. Change that surface. How you make the change stays on [agent experience optimization](https://www.agentrail.co/blog/agent-experience-optimization). Next window, check whether that same gap got smaller for the same kind of task. One gap a week is enough to see which change moved the count.