Flawed ground-truth data in agent benchmark "Humanity's Last Exam" underscores the need for better ways to measure agent effectiveness.
Primarily we want to measure our agents' velocity of output. We'll adapt from StarCraft 2 the idea of actions per minute (APM).
We demo the stats pane we added to OpenAgents which analyzes historical Claude Code conversations and shows your APM over the last hour, 6 hours, day, week, etc.
From this author's 30-day Claude Code usage across 277 sessions, we establish our baseline of 2.3 APM.
Rookie numbers! Now to pump them up...