Fleet: what each agent is doing now
In the app: Agents, then the Live state tab.
The Fleet page shows every agent as a card with its state right now, and what is being spent right now. Every state is derived from records, and the card shows the rule and the evidence beside it. An agent never reports its own state here, and Provantix reads records: it does not answer a waiting agent, start, stop or control anything.
In the product
Product capture, 15–16 November 2026. Open the image for full size. Counts and dates are examples from the captured workspace.
Read the state together with its explanation and last-seen time. This screen observes delivered records; it does not start or stop agents. A cost rate with unpriced calls is a lower bound.
The states
| State | The rule |
|---|---|
| Waiting on a person | The agent's latest enforcement verdict in the last 24 hours is ask (a tool asked a person to decide), and the agent has recorded nothing since. The card says what was asked, by whom and since when. |
| Failing | Its most recent three or more outcome-bearing calls all failed, the latest within 30 minutes. |
| Working | It recorded something in the last 5 minutes. |
| Idle | It recorded something in the last 24 hours, but not in the last 5 minutes. |
| Quiet | Nothing in the last 24 hours, or never seen. |
| Paused, retired, proposed | The agent's lifecycle. An agent that is not running is shown as what it is, never as quiet. |
They are decided in that order, so an agent that is waiting is shown as waiting even if it also has failures. A verdict or a decision record is not the agent doing work, so it does not make an agent "working".
What a card shows
The last step (the tool or action, the system, and when; click it to open the run), the calls in the last hour, the estimated cost today with the number of unpriced calls beside it (and "not recorded" when there is none, not zero), the open findings with the highest severity, and the owner. Cards that need attention come first. The state counts at the top filter the board.
The burn rate
Above the cards: the estimated cost per minute and the output tokens per second over the last 5 minutes, and the last hour as a bar per minute. It is the same per-record estimate as all cost, priced at ingest, not a billing figure. Priced calls are summed and the model calls with no price are counted beside them, so when any are unpriced the rate is at least what is shown. Records are placed by when they were observed, so a record that arrives late fills in a minute that has already passed.
What this does not do
- It does not know what an agent is thinking or waiting for beyond what a record says. "Waiting" comes only from an enforcement tool's recorded ask.
- It does not answer a waiting agent. The enforcement tool and the person do that.
- It does not show agents Provantix has no records for. See Coverage.
