Transcripts and analytics

Once your agent is taking calls, this is where you find out how it is doing. Every call is transcribed and scored, so you can read what happened and measure whether it is getting better.

Transcripts

The Transcripts page lists your agent's calls. Open one to read the full conversation, turn by turn — what the caller said, what the agent said, and the tools it called along the way.

Use transcripts to:

• Find the failure. When a call goes wrong, read it back to see exactly where — a misread intent, a wrong tool, a clumsy hand-off.
• Spot patterns. The same stumble across several calls is a prompt or tool fix waiting to happen.
• Confirm a fix. After you change the agent, watch new transcripts to check the behaviour actually changed.

Reading a call

A transcript shows the conversation in order, including the tool calls the agent made and what came back. That is usually enough to diagnose "why did it say that" — the model's behaviour follows from the prompt, the tools, and what the caller said, all of which are right there in the call.

Metrics

Beyond individual transcripts, the agent's overview surfaces recent-call signals at a glance, and the evals dashboards carry the voice-quality metrics that decide whether calls feel right:

• Latency and time-to-first-byte — how fast the agent starts responding.
• Interruptions and double-speak — talking over each other.
• Filler rate — how often fillers mask a wait.

Read transcripts to understand a single call; watch the metrics to know whether the agent is improving across many.

Closing the loop

Transcripts and metrics are the input to the next round of work: find the weakest behaviour, write an eval for it, fix the prompt or a tool, and publish. Reliability is built one fixed failure at a time.