Last updated: 2026-08-31 · 2 min read
Review results and improve an Agent
After deployment, review individual outputs separately from longer-term trends. Never change a live Agent in place: create a new draft from an improvement suggestion, evaluate it again, and deploy the approved version.
| Screen | What to inspect | What to do next |
|---|---|---|
| 実行レビュー (Run Review) | One run, its output, sources, and errors | Investigate a specific result |
| インサイト (Insights) | Quality, success, and execution trends for a Room | Identify an emerging pattern |
| 改善キュー (Improvement Queue) | System-detected improvement suggestions | Create a draft for review |
Start an improvement draft
- Read the evidence and scope of an improvement suggestion.
- Select 下書きを作成 (Create draft) when you want to adopt it.
- Check and refine the change in ジャーニー (Journey).
- Re-run シミュレーション (Simulation) with the previous cases and the affected case.
- Approve and deploy the new version only when it is ready.
Improvement Queue does not automatically fix or publish an Agent. A person must verify the proposal and its permission impact.
Related articles
- Evaluate, approve, and deployCheck expected results in simulations before deploying an Agent to a Room
- Run an Agent: manual, schedule, or webhookChoose a safe trigger for a manual run, recurring work, or an external event
- Troubleshooting Agent StudioCheck goals, evaluation, deployment, and run results when an Agent does not work as expected