An item nobody closes is worse than no item
Backlogs accumulate until nobody reads them, and an unread backlog is not a plan. Somewhere in nine hundred items are the thirty that still matter. The cost of finding them is why nobody looks.
Staleness is not age. A two-year-old item that describes a problem that still occurs is live. A three-month-old item that describes a screen that no longer exists is not. Sorting by date finds neither.
Most cases can be decided on three questions: has the problem still occurred, has a released item already solved it, and does another item already cover it? Each question must be checked against the current product, not against the item's own text.
The agent proposes dispositions with reasons. It closes nothing and merges nothing, because a closed item is one nobody finds again, and that decision belongs to the team.
How the agent is built in Agent Studio
Staleness rules live in prompt skills: what your team counts as resolved by other work, which areas were rewritten and when, and which items are deliberately left open.
The system prompt requires each proposed disposition to cite what makes it true: the release that solved it, the duplicate item, or the area that no longer exists. If the agent cannot cite that evidence, it leaves the item alone.
It forbids a disposition based on age or inactivity alone, which is the heuristic that quietly discards real work.
It requires duplicates to be proposed as a pair, with both texts shown, so the person deciding can see exactly what would be lost in a merge.
A Start Scheduled flow reads the backlog in batches through a Jira or Linear node. The Agent Chat node receives each batch with the latest release notes and the current area list attached.
Proposals are written in Google Sheets, one row per item, with its reason and evidence, a format a team can work through in a single session.
A Human in the Loop node holds every disposition. The agent produces a reviewed list. The team spends an hour instead of a quarter.
What this agent is built from
- Prompt skills: what counts as resolved by other work, which areas were rewritten, and what stays open deliberately.
- System prompt: every disposition cites what makes it true, or the item is left alone.
- System prompt: no disposition on age or inactivity alone.
- System prompt: duplicates proposed as a pair, with both texts shown.
- Jira (Integration): reads the backlog in batches. Linear and Monday fit the same slot.
- Google Sheets (Integration): one row per item, with its reason and evidence.
- When an ActionFlow is enough: If you only need items that have not been touched beyond a threshold, with their linked releases, build the workflow. That list is enough for an annual sweep.
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.