What should an approval gate for AI agents actually bind to? Lessons from the comments

My last post was about human approval for destructive MCP tools. The comments taught me more than the...