Production scores

Pre-release testing proves your agent is ready. Live monitoring proves it stays good with real users. Reached under Evaluate → Live monitoring, it continuously derives quality signals from the real conversations your agent handles after launch.

The metrics

• No-response / abandonment rate — how often users leave without getting a useful answer. Flagged when it climbs past roughly 10%.
• Tool errors — how often the agent's tools fail during real use.
• Empty retrieval — for knowledge agents, how often a search returns nothing useful. A high rate (above ~50%) points to gaps in your knowledge base.
• User frustration rate — an LLM judge estimates how often users are visibly frustrated (watch above ~15%).
• Latency (p50 / p95) — typical and worst-case response times, including how many model calls each turn takes.

Metrics carry color-coded health badges, a trend over your selected window (7 / 14 / 30 days), and a model-usage breakdown.

Drilling in

Every metric card is clickable: you drill straight into the real conversations behind the number and read the exact turns that triggered it. Open the dashboard, pick a period, scan the badges for anything red, and click into the offending conversations to see concrete examples.

What you get

An always-on quality signal on live traffic, with real examples rather than just aggregates — closing the loop between what you tested and what users actually experience. Findings here tell you what to add to your datasets and what to optimize next.