Skip to main content
Voice agents fail in ways text agents do not: a prompt that reads perfectly can produce an agent that talks over people, mishears digits, or refuses to hang up. Test on audio, not on paper.

​
The four stages

1

Test in the browser

Fastest loop for prompt changes — no telephony, no cost per attempt. Use the Web Call SDK or the test call in the dashboard. Good for wording; useless for audio quality.
2

Call yourself on a real phone

Everything that matters — carrier audio, endpointing under real latency, background noise, DTMF — only appears on a real call. Do this on the same telephony provider you will use in production.
3

Run the awkward-case list

See below. These are the cases that generate support tickets in week one.
4

Dry-run a small batch

Send a batch of 20–50 calls to internal numbers or consenting testers. Batch behavior — scheduling, guardrails, retries, concurrency — is not exercised by single calls.

​
The awkward-case list

Keep the list as a fixed script and re-run it after every prompt or model change. Regression, not first-run behavior, is what breaks live agents.

​
What to read after each test call

Open the execution in Call history and check four things:
  1. Transcript — did the transcriber hear correctly? Mistakes here explain most odd replies.
  2. Recording — how did the pacing actually feel? Read time is not talk time.
  3. Latency metrics — where did the time go, per turn — Read latency metrics for a call.
  4. Extracted data — did the fields your business depends on come out populated and correctly typed?

​
Testing graph agents and workflows

Graph agents have their own testing, validation and debugging tools — use them to verify every transition is reachable before making any calls. Validation catches structural faults; only real calls catch conversational ones.

​
Next steps

Production readiness checklist

Everything else before launch

Failures and retries at scale

What happens when calls fail

Plan concurrency

Before the first big campaign