An agent that waits before it acts
The defining habit of an agent on Czar is that it stops at the step that sends, spends, or changes a record and hands the decision to a person.
· 4 min read
IN SHORT
An agent on Czar reads and drafts on its own, but the step that sends, spends, or changes a record waits for a person. The reviewer sees the draft and the passages it was built from, then approves or stops. The run continues on that decision, and the decision is written into the record with the reviewer's name.
Most of what an agent does is harmless on its own. Reading a packet changes nothing. Looking up passages in the site's corpus changes nothing. Producing a draft changes nothing, as long as the draft stays a draft. The moment that matters is the one where the agent would send a message, spend money, or alter a record somewhere. On Czar, that moment waits.
Where the pause sits
Picture the run as a line. A packet arrives through the front door the desk already has. The agent reads what it may read, on the appliance. It drafts from those passages. Then it reaches the step that would act on the world, and it stops. A reviewer looks at the draft and what it used. The reviewer approves or stops the run. If approved, the allowed call is made. If stopped, the run ends there.
The pause is positioned exactly at the boundary between thinking and doing. Everything before it is reversible by definition, because nothing has left the appliance. Everything after it is the thing the site cared enough about to put a person in front of.
What the reviewer is given
A pause is only useful if the person at it can decide quickly and well. The reviewer on Czar is not handed a bare output and asked to trust it. They see the draft the agent produced, the passages from the corpus that the agent drew on, and the call the agent is proposing to make. The inputs that started the run are there too.
This matters because a reviewer's question is rarely "is this text acceptable" in the abstract. It is "did the agent read the right thing, and does the draft follow from it." Showing the passages answers that directly. A reviewer who sees that the agent drafted a denial from a policy clause that does not apply can stop the run in a moment, and the stop is recorded.
The decision becomes part of the record
When the reviewer approves or stops, that act is written into the run record alongside the inputs, the draft, and the call. Inputs, draft, decision, and call are one sequence someone can open tomorrow. The decision is attributed to the reviewer role and the person holding it.
This gives the site something it rarely has with automated work: a clean line of responsibility. The agent drafted. A named person approved. The call was made. If the outcome is later questioned, the record shows who looked at what and what they decided, without anyone having to remember.
Why this is not the same as supervising everything
A reasonable worry is that an agent that waits is an agent that needs a babysitter. It does not. The reviewer is not watching the run. They are called to one step, shown a draft and its sources, and asked for one decision. The reading and drafting happened without them. The call will happen without them once they approve.
The cost to the site is one person's attention at one point in each run. For a desk that already has a sign-off step, this often replaces a manual review of raw material with a review of a prepared draft and its evidence. The reviewer's time goes to judgment rather than to assembly.
Placing the step is a design decision
The operator decides where the approval step sits when they draw the agent in Agency Studio. For a desk whose only consequential action is at the end, there is one pause near the end. For a desk where the agent might take an intermediate action, such as requesting a document from another system, the operator can place a pause there too. The placement is frozen into the published version, so it cannot drift between runs.
The guide Placing the approval step covers how operators think about where the pause belongs. Approvals by named roles covers who may sit on it.
An agent on Czar does not act on the world without a person having said yes to that specific act, with the draft and its sources in front of them, in a way that is written down. That is the habit. It is why a site can let an agent read and draft freely: the step that could cause harm is the one step that always waits. The governance page describes how this fits with the rest of the platform's controls.
Questions this raises
- Can the waiting step be removed for low-risk work?
- The operator decides where approval is required when drawing the agent, and that placement is frozen into the version. Reading and drafting steps do not need to wait. But a step that sends, spends, or changes a record is the kind of step the platform expects a person to sit on.
- What does the reviewer actually see?
- The draft the agent produced and the passages it drew from, side by side, along with what the agent is proposing to do next. The reviewer does not have to reconstruct the reasoning. They approve or stop, and that decision goes on the record.
- What happens if no reviewer is available?
- The run waits. Nothing is sent, spent, or changed until a person decides. The operator can see runs that are waiting and the admin can decide who holds the reviewer role for that agent.
KEEP READING