Old is not the same as wrong
Refresh programs sort by publication date, because that date is easy to sort. The outcome is foreseeable. Evergreen explainers are rewritten for no reason, while a pricing page with out-of-date figures stays untouched because a typo on it was corrected last month.
Staleness applies to the content, not to the timestamps. A page that covers a protocol that has not changed is not stale even if it is three years old. A page that gives competitor pricing becomes stale within weeks.
Most refresh programs skip the judgment about leaving a page as it is. A page that ranks well and converts well is a risk to edit. Sometimes the right action is to do nothing.
An agent can do all three at once. A date sort holds none of them.
How the agent is built in Agent Studio
Staleness is domain-specific. The system prompt specifies what it means for you. It lists the page types that must never be rewritten without human approval first: pricing pages, legal pages, and any page that contains a regulated claim.
Prompt skills include your content standards and a list of facts that change often enough to need checking: your own pricing, the number of integrations you have, and your position relative to your competitors.
Built-in tools and MCP connections let the agent access the pages and their performance figures. Without performance data, a staleness judgment produces a list sorted by principle rather than by value.
The agent outputs a prioritized queue, with a reason for each page, and a section of pages it recommends not touching. Read the leave-alone section first when you check whether you trust the agent.
It suggests. It never edits. A person must approve before any change is deployed.
An ActionFlow can run the sweep on a schedule through Agent Chat, and write the queue into Notion or Linear so the review runs without anyone starting it. Because sessions last, an editor can then work through the queue in conversation and ask why a page was ranked where it was.
What this agent is built from
- System prompt: what staleness means in your domain, and the page types it may never touch unreviewed.
- Prompt skills: your content standards and the facts that change often enough to need checking.
- MCP connections: access to the published pages and their performance data.
- Agent Chat (Util): runs the sweep on a schedule from inside a flow.
- Linear (Integration): receives the prioritized queue as tracked work. Notion fits the same slot.
- Durable sessions: an editor can work the queue conversationally and challenge the ranking.
- When an ActionFlow is enough: If you refresh on a fixed annual cycle regardless of content, a scheduled workflow does it. The agent is for deciding what deserves attention.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.