How a call runs
Each agent reply passes through the same stages. Knowing them tells you which setting to change when a call goes wrong.
Getting audio to the engine
The zlicx voice engine receives calls over SIP from your carrier, or over WebRTC from the browser for preview and the website widget. Everything after this point works the same way for both.
Hear
The engine streams audio to the speech-to-text model you chose. The model returns words as they're spoken, and signals when the caller seems to have finished. If that signal comes too early, the agent talks over people who were only pausing.
Decide
The finished turn goes to your language model with the agent's instructions and the conversation so far. The engine reads the reply as a stream, so it can act on the first words straight away. It starts drafting while it waits for the end of the turn, and throws the draft away if the caller keeps talking. Discarded drafts aren't billed.
Act
The model can ask to use a tool: an HTTP request to your API, a tool on an MCP server you connected, or a knowledge base search. The result goes back to the model, which carries on with its reply. A slow tool means silence on the line, so have the agent say it's checking first.
Speak
The reply goes to text-to-speech in small pieces, so the caller hears the opening while the rest is still being written. If the caller interrupts, playback stops, and the conversation history keeps only the sentences they actually heard.
Record
When the call ends, zlicx writes the call log:
- The transcript, and the recording if you turned it on
- The analysis fields and evaluations you defined
- Usage in tokens, speech seconds and characters, which your wallet is charged for
- Your post-call webhook, if you set one
Where to look first
| What you notice | Stage | Try this |
|---|---|---|
| Cuts people off | Hear | Test another speech model in that language |
| Gets names or numbers wrong | Hear | Check the language setting, and read important details back |
| Answers wrongly with confidence | Decide | Narrow the instructions, and fix the knowledge base |
| Says done when nothing changed | Act | Add a prompt rule, and make the tool return clear errors |
| Long pause before replying | Decide or Act | Use a faster model, shorten the prompt, or speed up the tool |
| Sounds wrong or robotic | Speak | Change the voice, and ask for short spoken sentences |
Set up a test run with the first agent guide, or see each route in the API reference.