Monitoring Retell agents in production

Watch a Retell agent on the traffic it is handling now, with a score beside each call and an alert when quality moves rather than when a call fails.

Retell's dashboard answers whether calls are happening. Call counts, disconnection reasons, latency, concurrency, cost. All of it is about the call as an event.

Monitoring an agent means something else: knowing today that the answers got worse, on calls that all connected and all ended normally.

What Retell already gives you

More than most platforms in this list, and it is worth using before adding anything.

`call_successful`	Scored against success criteria you write. Your standard, Retell's model.
`user_sentiment`	The caller's overall sentiment for the call.
`call_summary`	A summary of what happened.
AI QA	A separate product that scores hallucination, knowledge base accuracy, latency and tool usage.

Retell AI QA is real and it is good at what it does. Two things about it shape where it fits: it evaluates a sampled set of calls, and it is metered per analyzed minute after the first hundred.

What watching live adds

Evidova scores the deterministic things on every call as it lands, in seconds. Latency, stalls, dead air, drops, the caller giving up.

The checks that cost a model call run on a share you set, with full coverage swept nightly. So the cheap signal is complete and the expensive one is budgeted.

Analysis fields are not on the call_ended payload. They arrive on call_analyzed, which is why a listener wired to the first one reads them as null.

Alerts that fire on quality, not on failure

Connect Retell in Setup and paste the webhook URL into your agent.
Retell signs its deliveries, so a request that fails the check is refused and recorded.
Pick the rules that matter enough to wake someone.
An alert on everything is an alert on nothing, and gets muted in a week.
Send them to Slack, WhatsApp or email.
The alert carries the conversation, so the first click is the evidence.

A bad release usually shows as a rule that starts failing on Tuesday and was fine on Monday, long before anyone files a complaint about it.