Whether the agent achieved the scenario's intended outcome

Agent fabricated information not grounded in tool responses

All expected tools were called with correct arguments

Accuracy of the agent's speech-to-text understanding

Response relevance to user queries

Consistent answers across the conversation

Agent correctly detected voicemail greeting

Count of times the agent interrupted the user

Milliseconds for agent to stop after user interruption

Response latency with P50/P90/P95/P99 percentiles

No connection drops, audio gaps, or timeout errors

Total silence duration in conversation

Agent avoided repeating the same information

Customer satisfaction score (0-100)

Overall conversation sentiment: positive/neutral/negative

Categorization of the call subject

Where in the conversation the user dropped off

Mean pitch frequency of agent voice

Ratio of agent speaking time to total time

Agent words per minute

Time before agent picks up

Agent did not produce unintelligible speech