AI
Human approval and audit in AI agents
Financial, public, destructive, credentialed and irreversible actions need an approval gate.
The prompt, tool arguments, model and version, policy decision and any override should all sit in tamper-evident logs.
