What's this about?
AgentAudit attaches to an agent's recorded trace and scores planning, memory, tool selection, invocation, correctness, alignment, security and execution integrity while attributing failures to stages.
What exactly is a “Composite Trust Score”?
It aggregates several dimensions beyond final task success. Two agents may finish equally often while one follows instructions and another succeeds through unsafe tool use.
Why is this worth noting?
Reported scores range from 95.1 for Claude Sonnet 5 to 22.6 for Gemini 2.5 Flash across nine tasks. A single fixed judge model — itself one evaluated model — is a significant acknowledged limitation.