AI agents can now be measured by how often they refuse to act

Researchers have created a new way to measure how AI agents behave when given tasks, focusing on how often they act versus how often they refuse. This means organizations can now see if an AI is more likely to follow instructions or say no, especially in risky situations.

Until now, it was hard to tell if an AI agent was genuinely being cautious or just failing to understand a command. This new measurement helps organizations understand the actual behavior of AI tools they deploy, not just their stated capabilities. It allows them to pick AI agents based on how they handle risk, rather than just how well they complete tasks.