Should we: * Allow the AI to make tool calls that perform assertions so they can be recorded? * Only allow AI to make assertions on itself an return true/false?
