Launch readiness thresholds
- Low risk: read-only, full trace, human gate, no sensitive data.
- Medium risk: draft writes, partial automation, reviewer approval, manual rollback.
- High risk: production writes, sensitive data, weak logs, unclear owner.
Static tool
Before an agent gets tools, score the risks that turn demos into production incidents.
The output should name launch blockers, required controls, approval gates, monitoring needs, and the incident-review template to use if the agent fails.
A launch blocker is a missing control that can let an agent take high-impact action without evidence, review, rollback, or a clear owner.
A high-risk score can be acceptable only when the workflow stays in a constrained pilot with strong human review, logs, and explicit non-production limits.
Novamente Weekly
Subscribe for scorecard updates, incident-review templates, and agent reliability notes.