Ship / Testing

Testing an agent

Three ways to try an agent, each answering a different question. Use them in order — the cheap ones catch logic bugs, and only a real call proves a transfer.

The ladder

Dry run — read the logic

In the builder. Type as the caller, see the reply, and watch what the agent did between turns — saved a variable, called an API, chose a branch. Runs the real engine over text, so routing and captures are exactly what a call would do.
Can’t tell you: anything about audio — pacing, barge-in, or how it sounds. Transfers are simulated (there is no phone line), and each send replays the conversation from the start, so an AI step can word things differently on a replay.

Test call — hear it

A voice call from your browser, no phone number needed. This is where you judge pacing, barge-in and pronunciation — the things a transcript can never show. Interrupt it on purpose. Pause mid-sentence. Say a proper noun and see if speech-to-text hears it.
Can’t tell you: whether transfers work — a browser caller has no phone leg to transfer.

Real call — trust it

Attach a number and dial in from a phone. The only test that exercises the whole path: carrier audio, caller ID, and transfers. Always do one of these before a customer does.

What to use for what

Does it capture what I need?Dry run — every save shows inline.

Does it take the right branch?Dry run — the path is highlighted on the canvas.

Does it sound human?Test call. A transcript can’t tell you this.

Can I interrupt it?Test call — try talking over it.

Does the transfer work?Real call from a phone. Nothing else proves it.

Does caller ID work?Real call from the phone you expect to be recognised.

Reading a call afterwards

Open any call under Calls. The transcript shows the conversation with the agent’s actions inline — API calls, saved variables, and each step of a transfer — so “why did it do that?” is usually answered by reading top to bottom. Failures appear in red where they happened.

🧪

Test the unhappy paths too. Say something off-topic. Refuse to answer a question. Go silent. Ask for a human immediately. Callers do all of these, and they’re where agents fall apart.

Before you point it at real customers

One real phone call, end to end, including a transfer.
Interrupt the agent mid-sentence and confirm it stops.
Try the flow with a caller who answers nothing useful.
Confirm the transfer destination is a line someone actually answers.
Check the numbers it reads aloud — see Voice. TKT-5001 should not be spelled out letter by letter.