Following a live run
A reviewer can watch a run as it happens: the packet arriving, the passages retrieved, the draft forming, the pause. Here is what that looks like and when it is worth doing.
· 4 min read
IN SHORT
A reviewer can follow a live run on Czar, seeing each step as it is written to the record: the packet that arrived, the passages the agent retrieved, the draft it produced, and the waiting step where a decision is needed. The view is the same record that will be opened later, growing in real time, and it is most useful in the first runs after a new version goes live.
Most runs on Czar are reviewed at the waiting step and never watched. But there are moments when a reviewer wants to see a run unfold: the first packets through a new version, an unusual kind of request, a desk that is still learning what the agent does. For those moments, a reviewer can follow a live run.
What following live shows
The run record is written as the run proceeds. Following live means watching that record grow. The reviewer sees the packet arrive through the front door. They see the reading step retrieve passages from the material the agent may read, and which passages. They see the drafting step produce its draft from those passages. They see the run reach the waiting step and pause.
At each stage, the record notes whether policy allowed the step. The reviewer following live sees those checks as they happen. If the agent reaches a point its version does not permit, the reviewer sees that the step did not occur and why.
The view is not a summary or a dashboard. It is the record itself, the same one that will be opened tomorrow, appearing in order. The run record and what it holds describes its contents.
When the run reaches a waiting step, the reviewer following live is already looking at everything they need: the inputs, the passages, the draft, and the proposed call. They approve or stop. The decision is written into the record with their name, and the run continues or ends.
A run waiting for a decision is visible to all reviewers for that agent, not only one who happened to be following. Whoever acts is the one named. Following live simply means the reviewer arrived at the pause with the context already in view, rather than opening it fresh.
When watching is worth it
Just after a new version goes live. The first runs under a new version are where design problems show. A reading step that retrieves the wrong material, a drafting step that misreads a clause, a waiting step placed too early. A reviewer watching the first few runs sees these before many packets have passed through. If something is wrong, they stop the run and the operator turns the address off and points it back. Pointing the address back at an earlier version covers that recovery.
When a packet is unusual. A desk sometimes receives a packet of a kind the agent has not handled often. A reviewer who knows it is coming can follow the run and see whether the agent reads and drafts sensibly for it.
When a desk is new to the agent. Reviewers who have not worked with an agent before often find that following a few runs live gives them a feel for what the agent does and does not do. After that, they review at the waiting step with confidence.
What following live is not
It is not supervision of the agent's reasoning. The reading and drafting steps proceed on the appliance without a person in the loop, because they change nothing outside it. The reviewer's decision point is the waiting step, and that is where they act. Following live gives context, not control over the intermediate steps.
It is also not required. The waiting step summons a reviewer whether or not anyone was watching. An agent whose runs are never followed live is exactly as safe as one whose runs always are, because the pause is the same. Following live is about catching design problems early and building confidence, not about safety at the step.
The operator's view
The operator for an agent can also see runs in progress, and in particular can see which runs are waiting for a decision. This tells the operator whether reviewers are keeping up and whether a version is producing more stops than expected. An operator who sees a cluster of stops on a new version has a reason to look at the design, and the records of those stopped runs show what the reviewers saw.
From live to finished
When a run ends, whether by an allowed call or a stop, the live view simply becomes the finished record. Nothing is lost in the transition. The reviewer who followed the run live and the colleague who opens it tomorrow are reading the same thing. Opening a finished run the next morning describes the later reading.
That continuity is the point. The appliance does not have one view for watching and another for auditing. It has the record, which can be read as it is written or after it is done. The governance page describes how the reviewer's view fits with the rest of the platform's controls.
Questions this raises
- Can a reviewer intervene in a live run before the waiting step?
- The reviewer's decision point is the waiting step. Reading and drafting proceed on the appliance without intervention, since they change nothing outside it. If a reviewer sees something wrong in the drafting, they stop the run at the waiting step. The operator can turn the address off to prevent further runs.
- Do reviewers have to watch every run live?
- No. The waiting step calls for a reviewer whether or not anyone was watching. Following live is a choice, useful when a version is new or a kind of packet is unusual. Most runs are reviewed at the waiting step and, if needed, opened later.
- Is the live view different from the finished record?
- No. It is the same record, seen as it is written. When the run ends, the view becomes the finished record. Nothing is shown live that is not kept, and nothing is kept that was not shown.
KEEP READING