Arena Experiments: Quality Check in Agent Mode

When you give a task in Agent Mode, Arena can turn your request into a short checklist of what you asked for. When the agent finishes, a separate model reviews the work and marks each item Completed, Incomplete, or Not Evaluated. The results show up in a Quality Check card under the agent's reply, with a count such as "3/4 completed."

What you can do

• Edit a requirement. Click the pencil next to an item to reword it. Press Enter to save, or click Revert changes to undo.

• Continue working. Send the agent back to finish incomplete items or to apply your edits. The follow-up appears in your chat labeled "Guided message."

• Dismiss. Close the checklist. Your next request starts a new one.

Your latest checklist also stays in the workspace side panel under Quality Check, so you can follow progress from turn to turn.

Not every message gets a checklist. Quick greetings and small talk are skipped, and the card appears only once the review is finished. Quality Check is available in some Agent Mode sessions, but not in coding sessions, and only in English for now.

Please note that this feature is currently part of an experiment, so it may not be available to all users.