Same words, different problem; different words, same problem
Three hundred requests might relate to no more than twenty different problems. The link between those problems is not a matter of vocabulary alone. A user who asks for a bulk edit and another who asks for a way to fix 40 records at once can be the same request. Two users who often ask for better reporting are not the same request.
To decide, you should weigh each person's situation rather than the features they mention, because their suggestions are shaped by what they already know about your product's capabilities.
Ticket counts can mislead. For example, 40 tickets from a single business account in one poor week are a weaker signal than 8 tickets from 8 different companies. A simple total does not show this.
The agent forms clusters from specific evidence and adds a note that names who is handling the request. Prioritization stays a choice for the product team, with the person who owns the roadmap.
How the agent is built in Agent Studio
Prompt skills hold product vocabulary and several cases you previously accepted and rejected. Those cases show the agent how fine a distinction your team wants when it treats problems as separate.
The system prompt requires member quotes in each cluster, plus the reason for joining the two items. It also requires an alternative interpretation whenever a join is uncertain.
The agent does not treat the count as evidence of priority. Distinct accounts, plan level, and recency are reported as separate facts. None of this is combined into a single score.
The flow collects tickets, reviews, and notes, then passes them with the existing cluster set to the Agent Chat node in batches. Clusters accumulate rather than being rebuilt each time. Each row is a cluster and its members. Linear receives only items from clusters that a person has accepted.
A Human in the Loop node sits between clustering and the tracker. The agent proposes a cluster. The product manager then accepts, splits, or rejects it.
What this agent is built from
- Prompt skills: feature names, customer vocabulary, and known synonym cases.
- Prompt skills: previously accepted and rejected clusters, which set your granularity.
- System prompt: member quotes and a stated reason on every join.
- System prompt: count, distinct accounts, plan, and recency reported separately, never scored.
- Agent Chat (Util): batches arrive with the existing cluster set, so clusters accumulate.
- Airtable (Integration): one row per cluster. Linear receives only the accepted ones.
- When an ActionFlow is enough: If you only need to group a few hundred items a month into themes with attached quotes, build the workflow.
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.