Use the agent to make refinement teach more — not to pre-write the answers.
The team friction Refinement is the team’s most expensive hour — the whole team and stakeholders in one room — and its real product is learning: shared understanding, problems chewed together, surprising solutions. Yet that hour gets eaten by low-value scanning — re-reading tickets, hunting duplicates and dependencies, spotting half-baked items. Teams then “fix” the cost by having the PO pre-specify everything alone, which guts the very learning the meeting exists for.
The pattern Use an agent to prepare refinement in service of learning, not instead of it. Its sweet spot is the hard scan: a long backlog where overlaps, patterns, and dependencies aren’t easy to spot by eye. Let it surface those so the team’s scarce attention goes to thinking together — then stop. A PO bringing an unknown, unspecified problem for the team to brainstorm isn’t a gap to automate away; it’s the point.
Why it works with AI Scanning a hundred tickets for patterns is tireless, cheap agent work and slow, inconsistent human work — so offloading it is pure gain. The judgement about which problems deserve the team’s collective head stays human. You spend the reclaimed time on learning, the one thing this meeting can’t get anywhere else.
Watch-outs Two traps. Silo — don’t let prep become a PO-only ritual; that’s local optimization that quietly rebuilds the handoff. Over-specification — an agent makes it easy to arrive with everything pinned down (often as a wall of text no one reads), killing the brainstorm and the surprises a team finds solving a problem live. Prep to open the conversation, not to close it.
Related Prototype, then bring it to refinement · Executable specs · Two people per story