Voice agent observability
Voice agent observability is the practice of instrumenting AI phone agents so operators can see what the system is doing in production: how many calls run, how they end, what fails, how long responses take, and which agent or configuration is responsible for each number.
An observability surface typically combines summary measures (call counts, success rates, failure counts, average duration), time-series views that reveal trends and spikes, per-agent breakdowns, and drill-down paths from an aggregate number to the individual calls behind it. The goal is that when a number moves, an operator can find out why without exporting data or guessing.
Observability sits between two neighboring practices. Call monitoring observes a single live conversation so someone can intervene in the moment. Call analytics asks business questions of accumulated call data. Observability is the operational layer in the middle: continuous, aggregate, and tied to the health of the system rather than to any one call or business report.
Why it matters for AI phone calls
AI agents fail differently from software services. A deployment can pass every health check while conversations quietly degrade — a prompt change makes the agent verbose, a model update shifts its judgment, a tool starts timing out mid-call. None of these throw errors; they show up as drifting success rates, rising handle times, or a growing share of callers who hang up early.
Attribution is the other hard part. Aggregate numbers blend infrastructure failures (which are the platform's fault) with conversation failures (which are usually the configuration's fault), and blend one struggling agent into a fleet average. Useful observability separates failure classes, excludes infrastructure errors from conversation-quality denominators, and lets operators compare agents and time windows directly.
Sampling deserves care. Busy systems often compute detailed conversation metrics from a sample of recent calls; a good surface says when it is sampling and how many calls the sample covers, so an operator does not mistake a partial view for the whole.
In practice on ThunderPhone
ThunderPhone's Observability page (under Quality & Testing in the dashboard) scopes everything by a time window and an agent filter: summary cards for calls, success rate (AI-graded, with infrastructure errors excluded from the denominator), early disconnects, errors, and average duration; volume and success-rate charts; a per-agent performance table; and a recent-failed-calls list with end reasons. On very busy windows detailed metrics are sampled from the most recent calls, with a notice saying how many calls the sample covers.