the model is maybe 2% of an agentic system. the harness (permission gates, tool routing, context compaction, recovery) does the other 98.
--dangerously-skip-permissions disables part of the 98 because we trust the model more than the scaffolding it runs inside.
that's backwards.