Agent

The agent is a chat over everything in your workspace. Ask it why clicks fell last week, which pages a deploy broke, or whether Googlebot is being blocked, and it looks the answer up in your data before it replies, with the numbers behind it. Open it from Agent in the sidebar.

What it can read

The agent reads the same data the dashboard shows, scoped to the workspace you are in:

• Search Console performance for Google, and Bing's when Bing is connected, by page, query, country and device, with comparisons to the period before.
• Opportunities, sitemaps, Google Analytics landing pages and traffic.
• Cloudflare crawler traffic, error paths and firewall rules.
• Deploys with their site checks, the latest site crawl and its findings, and the last digest.
• The index status of one of your pages, from Search Console's URL Inspection.
• Reports shared with the workspace, their comments and the files the team uploaded.
• The team, connections, webhooks and their deliveries, request logs, activity, and usage.

It reads aggregated results, not raw tables, and it only ever sees workspaces you belong to.

Workspace context

Settings → Context holds a document about your site that the agent reads at the start of every conversation: what the site is, its markets and languages, URL templates, competitors, past migrations and incidents, and any rules the team wants every analysis to respect. When you tell the agent a lasting fact and ask it to remember it, it saves it there. Every save is a new version that the team can see and roll back.

What it does, and what it proposes

Most of what the agent does is reading. A few actions change something:

• Proposed for your approval: starting a site crawl, which is metered, and sending a webhook delivery again. The agent shows the proposed action with Approve and Cancel, and nothing happens until you approve it.
• Done when you ask: saving a fact to the workspace context, filing a report request for the team after you agree to it, and creating a share link for a report, which only a workspace owner or admin can do.

Everything the agent proposed or did is listed under Settings → Agent as the task history.

If you ask for something the agent cannot do, it says so rather than improvising, and the request is logged for us as a gap to build.

Page text is data, not instructions

Some of what the agent reads was written by someone other than you: the content of web pages, titles, third-party text. Any of it could be phrased as an instruction. The agent treats all of it as quoted material to read and summarise. It never follows instructions found inside it, never lets it choose what to do next, and tells you when such text asks it to act. The workspace context is handled the same way: it informs the analysis but never tells the agent to do anything.

Cost

Agent conversations are paid from the account's AI balance. Settings → AI Usage shows what is left. When the balance cannot cover a conversation, the agent does not run until it is topped up.