Token consumption guards: large-read warning (100KB+), read budget limit, subagent cap, and token budget tracking. Reduce waste by 20-40%.