Dashboard
Activity, answer quality, and anything waiting on you.
The Dashboard is the workspace's home page and where sign-in lands. It answers product questions — how much are agents used, how good are the answers, is the knowledge base healthy, is anything waiting on a human — and links out to Observability and Jobs for the deep dive.
It needs trace:read, like the traces it is built from. Owners and admins
hold it; anyone without it (a plain member, a custom role) is sent to Agents
instead.
Controls
Two controls reframe the page together: an agent filter (empty means the whole workspace) and a date range, which defaults to the last 24 hours. Both live in the URL, so a view is a link you can send. Three cards ignore both, because their answer doesn't depend on a window: Needs attention, Knowledge base health, and Recent conversations. Agents by activity follows the range but not the agent filter.
KPIs
| Tile | What it counts |
|---|---|
| Conversations | threads started in the range |
| Messages | messages in those threads |
| People talking to agents | distinct people who started them |
| Positive feedback | share of thumbs-up ratings |
Each compares against the preceding window of the same length, and says "no prior data" when that window had nothing. A strip below adds agent runs, tokens, p95 latency, and error rate.
Attention items
A one-line strip at the top counts everything waiting on someone and jumps to the full list at the bottom. Each group shows its true count and up to five items:
- Tool calls awaiting approval — flagged when an app trigger started the thread, since nobody has it open and it can wait forever.
- Failed jobs from the background queue.
- Knowledge sources that failed to process.
- Sources needing reindex — the folder's parsing or chunking settings changed since they were indexed.
- Connected apps needing reconnection.
- Triggers reporting errors.
When nothing is waiting, the strip just says so.
Answer quality
Evaluation tiles headline the scores you care about, read from
evaluation results for the page's range and agent. A tile
shows the rate for a categorical evaluation or the average for a numeric one;
results whose label scores null are counted but left out. Scores are filed
by when the traffic happened, not when it was scored, so a sweep over last
month lands in last month. A "Score over time" chart draws one line per
evaluation.
Customize picks up to four evaluations; the choice is saved in the org's
system settings, so saving needs setting:update. Until someone picks, the
first evaluations that exist are shown. Unticking everything hides the tiles,
a deleted evaluation's tile drops off, and a workspace with no evaluations has
no tiles at all.
The other cards
- Conversations started — over time, split by where each came from: chat, schedule, or app trigger.
- Response quality — thumbs up and down, the positive rate, and the reasons people gave when unhappy.
- Agents by activity — the top agents by conversations, with messages, people, and ratings.
- Recent conversations — the latest threads, as a live feed.
- Knowledge base health — sources by ingestion status across folders.
- Knowledge coverage — how many knowledge searches returned at least one passage, and the questions that didn't, grouped: No match (the content is missing) or Not relevant (passages came back but none answered). Searches that errored are left out of the rate, so an outage doesn't read as a content gap.
The same numbers are available over the API under /dashboard/*; see the
API reference.