Add a Site agent to assess and update your site
A Site agent is a Hub-assisted Operations agent assigned to one of your Sites. It can assess the site, save prioritised recommendations, turn a recommendation into a reviewable page or navigation proposal, and — when configured — apply changes automatically on a schedule. ## Before you begin - You must be a workspace owner or administrator. - The Site must already exist in your workspace. - Decide whether you want assessments only, reviewable proposals, or automatic updates. ## Create a Site agent 1. Open **Sites** in your workspace. 2. Select the site you want to configure. 3. Open the **Agent** tab. 4. Select **Set up agent**. 5. Write instructions for the agent — for example, the site’s purpose, audience and what kinds of improvements to look for. 6. Configure its schedule using the **Schedule** settings. 7. Save the agent. The agent editor reuses the standard Operations agent interface. Instructions and scheduling options are the same. ## Run an assessment Select **Assess site** at any time to run a read-only assessment of the site. Assessments: - Are always read-only, regardless of the agent’s other settings. - Check the site’s pages against the agent’s instructions and any configured evidence sources. - Save a brief with prioritised recommendations and explicit coverage. - Validate page references against the inspected page list. - Cannot invoke publishing, data mutation or GitHub tools. The assessment brief appears in the **Agent** tab with a summary, recommendations and evidence coverage. ## Make an immediate collection-based request You can send the agent a one-off request without changing its standing instructions or schedule: 1. In the **Agent** tab, find the **Request** composer. 2. Type your request — for example, *Prepare a Tech Stack page from the Technology collection*. 3. Select which collections the agent may read for this request. 4. Select **Start run**. The original request appears alongside its run results in the Activity section. Stable request IDs deduplicate retries, so resubmitting the same request does not create duplicate runs. ## Review saved recommendations After an assessment, recommendations appear in the **Agent** tab ordered by priority. Each recommendation shows: - What the agent observed. - What change it suggests. - Supporting evidence. Expand a recommendation to read its full brief. Select **Prepare change** to turn it into a reviewable proposal. ## Turn a recommendation into a proposal 1. Find the recommendation you want to action. 2. Select **Prepare change**. 3. Hub queues a preparation run — one run per recommendation is created; repeated requests reuse the same run. 4. When the run completes, the proposal appears in the run result with exact before/after fields. 5. Review the diff, then select **Apply** to apply the change or **Dismiss** to discard it. Version checks protect against intervening edits — if the target page was changed after the proposal was created, Hub rejects the apply and asks you to review the current content first. ## Configure the run mode A compact settings disclosure in the **Agent** tab controls what the agent is allowed to do on scheduled and automation runs: | Mode | What scheduled runs do | |---|---| | **Assess only** | Save assessments and recommendations. No changes are prepared or applied. | | **Prepare for review** | Save assessments and prepare reviewable proposals. No changes are applied automatically. | | **Apply automatically** | Save assessments, prepare proposals and apply approved changes to the live site. Requires explicit administrator consent. | **Assess site** always forces read-only mode regardless of the configured setting. Automatic mode explicitly enables live changes for the assigned site and can use separately permitted Data collection writes. ## Configure evidence sources Site agents can use external evidence when assessing your site: - **Public competitor or market pages** — enter URLs to include as reference material. Reports record dated source coverage and compare previous excerpts where available. - **Linked form submissions** — select a connected form. The agent samples recent responses (up to 100 per form) to understand visitor intent or common questions. Configure sources in the **Sources** disclosure within the Agent settings. Limits are disclosed: public HTML excerpts, first 100 site pages for linked-form discovery, explicit sample and text caps. ## Draft replies from form submissions When form submissions are configured as a source, the agent can create pending reply drafts in the existing record draft workflow: - Draft writes recheck permissions and submission identity in a transaction. - Drafts are deduplicated per response — one draft per submission. - Subsequent human edits to a draft are preserved and not overwritten. - No reply is sent automatically. Review and send reply drafts from the form’s **Responses** area. ## Review results Run results, proposals and recommendations all appear in the **Agent** tab: - The **Activity** section lists recent runs with their status and a compact summary. - Expand a run entry to see the full brief, any proposals and evidence coverage. - Select **View run history** to open the unified run history filtered to this agent. ## Shared blocks Site agents cannot apply changes to shared blocks. Shared blocks always require human review regardless of the configured run mode.