Last updated: 2026-08-31 · 2 min read
Evaluate, approve, and deploy
Creating a design does not run an Agent. First evaluate it against representative cases, then have a person approve the version that may be deployed.
| Stage | Who acts | Why it matters |
|---|---|---|
| Draft | You | Confirm the intended job and boundaries |
| Simulation | Terasu and you | Evaluate results against representative inputs |
| Approval | You | Choose the version that may be deployed |
| Deployment | You | Choose the Room and run method |
Run a simulation
- Open シミュレーション (Simulation) from the sidebar.
- Confirm the draft and evaluation cases.
- Run the evaluation and read the result and its reasons.
- Return to ジャーニー (Journey) to adjust the goal or conditions when a result is insufficient.
- Approve only the version that meets the intended outcome.
Review whether the result is useful, grounded in the intended information, and free of actions you prohibited. Include realistic meeting notes and common exceptions in evaluation cases.
Approval is not automatic execution
Approval means that a version can be deployed. Deployment then chooses the Room and run method. A schedule or an enabled webhook can start future runs automatically; that is a separate decision from approving the Agent definition.
Before deployment, confirm the Room, permitted connections, write or delivery boundaries, and at least one manual run.
Related articles
- Create your first read-only AgentUse plain language to create a safe first Agent that organizes meeting notes without changing anything
- Run an Agent: manual, schedule, or webhookChoose a safe trigger for a manual run, recurring work, or an external event
- Review results and improve an AgentUse Run Review, Insights, and Improvement Queue for different parts of the improvement cycle