On this page
A voice agent looks like one product from the outside. Inside, it is five jobs that run one after another, several times a minute, while a person waits on the line. Knowing those jobs helps you judge what will go wrong in your own calls, and which setting to change when it does.
1. Hear
Phone audio reaches our voice engine over a SIP connection, or through the browser if you are using preview or the website widget. The engine streams it to the speech-to-text provider you picked for that agent. The provider returns words as they are spoken and also signals when it thinks the speaker has finished.
That second signal is harder than the transcription. Stop too early and the agent talks over someone who was only pausing. Stop too late and the line goes quiet.
2. Decide
Once the turn is complete, the transcript goes to the language model along with your prompt and the conversation so far. On our engine that can be a model from Anthropic, OpenAI or Azure, or one reached through OpenRouter. We read the reply as a stream, so we can act on the first words before the last ones exist.
3. Act
If the model decides it needs something, it asks for a tool. A tool might be an HTTP request to your order system, a search of your knowledge base, or a tool on an MCP server you connected. There is also a built-in tool for ending the call. The result goes back to the model, and it continues the reply.
Tools are where an agent stops being a talking FAQ. They are also where most mistakes become expensive, which is why we cover them separately in connecting an agent to your systems.
4. Speak
The reply is cut into sentences and sent to text-to-speech one at a time, so the caller starts hearing the first sentence while later ones are still being written.
If the caller interrupts, the engine stops playback. It then keeps only the sentences the caller actually heard in the conversation history. Without that step, the model would believe it had said things nobody heard, and its next answer could refer to them.
5. Record
When the call ends, your record is built: the transcript, the recording if you enabled it, the structured fields you asked the analysis to extract, and any evaluations you defined. A post-call webhook can send the result to your own system. Usage is reported turn by turn in tokens, speech seconds and characters, and that is what your wallet is charged against.
What this means when you are testing
Sort each problem you hear by the job it belongs to:
| What you notice | Where to look first |
|---|---|
| The agent cuts people off | Hear: turn detection and speech model |
| Names and numbers come back wrong | Hear: speech model and language setting |
| It answers confidently and incorrectly | Decide: the prompt and knowledge base |
| It says "done" when nothing changed | Act: tool design and the prompt rule |
| Long silence before replying | Measure each stage; see why replies take the time they do |
Pick one call type and walk through it with the call test kit. If you want to go through your own flow with us, get in touch. The operations workflow page has more on which calls suit this approach.
See which calls suit this.
Choose a call with a clear result and systems your agent can reach.