Rollout delle azioni-agente

Govern how much your AI agents may act on their own. Climb the autonomy ladder (advisory → canary → percentage → full) class by class, arm the guardrails that pause them when a metric goes bad, promote a proven policy from staging to production, and use the global kill-switch to stop everything in one click. Every action is a write: it requires your confirmation, passes through the safety wall, and is audited and reversible.