AWS introduces AEM, a turn-level metric to isolate cascading agent errors

AWS ML Blog · rss · 2026-09-10

AWS ML Blog introduces the Agent Evaluation Metric (AEM), a decomposable, turn-level measure of multi-turn agent quality, starting with correctness.

Problem: one early mistake cascades through later turns; outcome-level evaluation can't tell root causes from downstream effects, and holistic scores can't distinguish factual errors from missing fields or wrong tool calls.

AEM approach:

Original post →

More from coding & agent

coding & agent channel →