Have an agent's work reviewed

Needs review before it counts: adding a second agent that checks another agent's work, what approved, sent back and not reviewed mean, and what a review can and cannot do to your run.

An agent can have its work checked by a second agent before it counts. The reviewer reads what the first agent produced, says whether it is acceptable and why, and either approves it or sends it back for one revision.

Turning it on

Open an agent's configuration and switch on Needs review before it counts. Three choices sit behind it:

Choice	When a review happens
Off	Never. This is the default.
Always	Every time this agent finishes.
Only when the agent is unsure	When something about the agent's own work looked uncertain — it said so itself with low confidence, its output did not match what it declared it would produce, or a check of its work rejected it and it had to be asked again.

Reviews work on every plan. One detail of Only when the agent is unsure does not: the first of its three signals — the agent saying so itself — is a note on the team conversation, which is part of Founding membership. On Free and Pro that signal is silent and the choice decides on the other two. Always and Off are the same on every plan.

Turning the toggle on adds a real reviewer card to your canvas, right after the agent it checks. You can see it, open it, read its instructions and delete it. Deleting the card is how you turn the review off again.

The reviewer card shows Reviews [the agent] in place of its own toggle. A reviewer is never itself reviewed.

You can also set a default for the whole workflow, so new agents get review without you switching each one on. It lives in Settings → Workflow, as the row Needs review before it counts (default for this workflow), with the same three choices minus Inherit — Off, Always, Only when the agent is unsure — and the line A card's own setting wins over this. underneath. An individual agent's own setting always wins over the workflow default.

In the chat builder you can just ask: "have the writer's work reviewed before it counts", and the same reviewer card is added.

What the reviewer can do

Very little, on purpose. A reviewer only reads. Its card is fixed to the read-only agent type and its tools are limited to reading, searching and listing — it has no write, no edit, no shell, no connected servers, no phone tools, no powers, no vault secrets, and no output file of its own. Cerevisor refuses to save a reviewer card that has been given any of those by hand, and narrows its tools again when the run actually starts.

Its job, written on the card for you to read and change, is to read what the producer said it would produce, post one review result addressed to that agent, attach at least one piece of evidence, and say why in a sentence.

What you see during the run

The review happens on the run's board, and the outcome shows on the agent's tile in the run window's Overview:

Shows as	Means
Under review / Waiting for its reviewer	The producer has finished and its reviewer is reading.
Reviewed	The reviewer approved the work.
Sent back	The reviewer did not accept it and said why.
Nobody checked this	A review was asked for, but nothing usable came back.

A review result that carries no evidence attached does not count. The board says so: This review result has no evidence attached, so it does not count. A reviewer that shows its working without ever saying whether the work was accepted is called out too. Silence is never read as approval — and it is never read as rejection either.

Only the agent that was actually asked to review can settle a review. Another agent posting a review result on the board — or you posting one from your own seat — does not overwrite the reviewer's answer.

What happens to the work

Approved — nothing changes. The work is passed on as it is.
Sent back, with evidence, the first time — the producer gets one revision. The reviewer's own sentence goes into its next turn, and one more review round follows.
Nothing usable came back the first time — the reviewer said nothing, or said nothing that counts. The work is marked Not reviewed: …, the run carries on, and there is no second round. A reviewer that stayed quiet can never stop your work, and nobody reads its silence as approval.
The second round settles nothing either — sent back again, or again with nothing that counts. Only here does the work get marked Sent back by its reviewer: … so everyone downstream reads why, and the run halts by default. The workflow's "halt on contract failure" setting is what decides that; turn it off and the marked work simply flows on instead.

The producer's work is never thrown away. Only its opening line is marked. Everything it wrote is still there, in the run window, in the export, and to the agents after it.

What a review cannot do

A review result is a note on the board, and the rule the board is built on holds here too: a note is information, never authority.

A reviewer cannot grant the agent it reviews any permission, tool, file or setting. It cannot widen anything at all.
The one thing a Sent back can do is stop a run you already started. It can never make a run do more.
Nothing a reviewer says is read by any permission check, approval prompt or safety wall.

Known limits

Reviews apply to a full workflow run. Re-running a single agent on its own, and a workflow invoked as a Power by another agent, do not pair reviewers.
The reviewer reads roughly the first 3,000 characters of the producer's work. On a long output, "approved" honestly means the opening looked right.
Setting a workflow-wide default of Always on a workflow whose agents have no reviewers blocks Run until reviewers are added or the default is changed in the Settings row above.
None of this has been exercised by hand in the running app for this release.