KnitKnot

# Review a proposed Agent version update

Compare a proposed page change with its current text, approve or reject it, and read the cycle record after publication.

## At a glance

  1. 1. **Open the queue**: In Content → Updates, choose a group and inspect its example change.
  2. 2. **Compare the wording**: Switch to Compare to see what comes out and what goes in, then check the evidence.
  3. 3. **Decide and follow up**: Approve or reject one proposal, or approve the displayed group count; open the cycle record to inspect the result.

If **Content → Updates** appears in your workspace, open it in the [console](https://app.knitknot.ai/content/updates). This workflow proposes changes to Agent versions of eligible pages. The page keeps the proposed text, approval decision, published version, and later measurement in one cycle record. Workspaces without managed updates go to **Content → Playbooks** instead.

Updates does not mean every page will be rewritten each week. A cycle can leave a page alone, defer it because of a cap, or hold a proposal that needs your decision.

## Review one proposed change

  1. 1. In **Awaiting your approval**, select a group or **View diff**. Groups collect proposals by change type and page class. The row gives you a page count and an example proposal.
  2. 2. Read the highlighted **Agent version** text. This is the proposed wording in page context. Check the evidence shown with each changed region.
  3. 3. Select **Compare** to see the original sentence beside its replacement. **Base only** shows the existing page text without the proposed additions.
  4. 4. If the proposed claim is about your product, check it against the [approved fact](/docs/review-and-approve-facts/) and its source. For example, when a revision changes a plan limit, confirm the exact limit and its date in the ledger before approving.
  5. 5. Select **Approve this one** or **Reject**. Rejecting leaves the served page as it is.

The default view emphasizes the proposed Agent version. Use **Compare** whenever a correction removes a sentence; the removed text appears there. If the current page text cannot be loaded, the review shows the proposed additions without inventing a base page. Wait for the source text or open the page record before approving a change you cannot assess.

## Approve a group with care

**Approve all** applies to the number of revisions shown for the selected group. The diff below the group is an example, not a separate review of every page in that group. Use individual approval when each page needs its own factual check.

A confirmation dialog repeats the count. If the group changes before the request reaches the server, the approval is refused and the row shows the reason. Refresh the queue and review the new count before trying again.

## Check the approval policy

Open **Approval policy** from Updates. Each page class starts in **Manual** mode. An **Auto** switch is available only after that class's evaluation gate is green on your workspace's own content. A switch change remains a draft until you select **Save policy**. If the server refuses it, the page shows the reason beside that class and restores the stored value.

Auto publication still creates a change record and a version that can be rolled back. Keep classes in Manual when your team must inspect every proposed statement before publication.

## Read what happened after the cycle

The Updates page shows the latest completed cycle's conclusion, the current cycle's page ledger, and a link to its **Full cycle record**. A running cycle has no conclusion yet. Open the record to see how many pages were planned, deferred, published, or needed attention. The page rows include check failures and approval information when those records exist.

**Measured movement** compares the benchmark runs around the cycle. A refused or incomplete measurement has no numeric result to interpret. When a holdout experiment ran, the record also shows treated and held-out pages separately. A change in visibility after publication is a measured association unless the experiment supports a stronger conclusion.

For customer-owned content work, start with [Issues and playbooks](/docs/issues-and-playbooks/). For the benchmark evidence behind a proposed correction, see [Investigate a claim in an AI answer](/docs/investigate-a-claim/).

Docs navigation
Docs › Guides

Review a proposed Agent version update

Compare a proposed page change with its current text, approve or reject it, and read the cycle record after publication.

Updated

At a glance

  1. Open the queue

    In Content → Updates, choose a group and inspect its example change.

  2. Compare the wording

    Switch to Compare to see what comes out and what goes in, then check the evidence.

  3. Decide and follow up

    Approve or reject one proposal, or approve the displayed group count; open the cycle record to inspect the result.

If Content → Updates appears in your workspace, open it in the console. This workflow proposes changes to Agent versions of eligible pages. The page keeps the proposed text, approval decision, published version, and later measurement in one cycle record. Workspaces without managed updates go to Content → Playbooks instead.

Updates does not mean every page will be rewritten each week. A cycle can leave a page alone, defer it because of a cap, or hold a proposal that needs your decision.

Review one proposed change

  1. In Awaiting your approval, select a group or View diff. Groups collect proposals by change type and page class. The row gives you a page count and an example proposal.
  2. Read the highlighted Agent version text. This is the proposed wording in page context. Check the evidence shown with each changed region.
  3. Select Compare to see the original sentence beside its replacement. Base only shows the existing page text without the proposed additions.
  4. If the proposed claim is about your product, check it against the approved fact and its source. For example, when a revision changes a plan limit, confirm the exact limit and its date in the ledger before approving.
  5. Select Approve this one or Reject. Rejecting leaves the served page as it is.

The default view emphasizes the proposed Agent version. Use Compare whenever a correction removes a sentence; the removed text appears there. If the current page text cannot be loaded, the review shows the proposed additions without inventing a base page. Wait for the source text or open the page record before approving a change you cannot assess.

Approve a group with care

Approve all applies to the number of revisions shown for the selected group. The diff below the group is an example, not a separate review of every page in that group. Use individual approval when each page needs its own factual check.

A confirmation dialog repeats the count. If the group changes before the request reaches the server, the approval is refused and the row shows the reason. Refresh the queue and review the new count before trying again.

Check the approval policy

Open Approval policy from Updates. Each page class starts in Manual mode. An Auto switch is available only after that class’s evaluation gate is green on your workspace’s own content. A switch change remains a draft until you select Save policy. If the server refuses it, the page shows the reason beside that class and restores the stored value.

Auto publication still creates a change record and a version that can be rolled back. Keep classes in Manual when your team must inspect every proposed statement before publication.

Read what happened after the cycle

The Updates page shows the latest completed cycle’s conclusion, the current cycle’s page ledger, and a link to its Full cycle record. A running cycle has no conclusion yet. Open the record to see how many pages were planned, deferred, published, or needed attention. The page rows include check failures and approval information when those records exist.

Measured movement compares the benchmark runs around the cycle. A refused or incomplete measurement has no numeric result to interpret. When a holdout experiment ran, the record also shows treated and held-out pages separately. A change in visibility after publication is a measured association unless the experiment supports a stronger conclusion.

For customer-owned content work, start with Issues and playbooks. For the benchmark evidence behind a proposed correction, see Investigate a claim in an AI answer.