Operate an agent product
Run every customer's agent from one view: what each did and for whom, its pass rate, its cost per Action, and one call that pauses them all.
Operating an agent product means watching every customer's Agent at once and stepping in on one. Each Action is on the record, graded and costed, so the numbers come from the work itself.
What to watch
| Number | Read from | Act when |
|---|---|---|
| Pass rate per task type | The Evaluator's verdicts | It falls below the rate the Offer was launched on |
| Pass rate per Tenant | The same verdicts, by for | One Customer's agent fails more than the rest |
| Approvals asked per Task | Each Customer's Approvals | Customers answer more than they expected |
| Cost per Action | Cost per action | Model and upstream cost eats the margin |
| Replies answered on time | The Role's standards | A seat misses its response time |
One Tenant failing while the rest pass is a fact about that Customer's work, not the product. Read its failed cases before changing the Role.
Follow one task
Every Action records who did it, for whom, through which channel and on which record. Traces follows one Task from the Message that started it to the verdict that billed it, and see what an agent did is the guide.
Step in
| To | Call | What happens |
|---|---|---|
| Stop one Customer's agent | Agent.pause with agents | It starts nothing new; running Actions finish |
| Stop every customer-facing agent | Agent.pause with customerFacing | Each writer gets a holding reply from the Role address |
| Start again | Agent.resume | It resumes with Memory, Mandates and Limits unchanged |
| Reverse one Action | Action.undo | Records it created leave every list and stay restorable; a sent Message stays sent |
A Customer pauses its own agent from its Portal the same way. Pause every customer-facing agent is the guide for the builder's stop.
Ship a fix
A fix is a new Version: edit the Role or its eval cases, and the gate replays every case before it ships. Add the failed case first, so the fix proves itself. Moving existing Customers onto it replays their own cases too: Agent upgrades, status planned.