Skip to main content
A Zeptar agent is a real-time voice assistant. You describe how it should behave in a system prompt, and the platform runs that behavior through a speech pipeline so people can just talk to it.

​
The conversation pipeline

Every spoken turn flows through four stages:
  1. Speech-to-text — the caller’s audio is transcribed.
  2. Language model — the transcript plus your system prompt produce the agent’s reply.
  3. Text-to-speech — the reply is synthesized back to audio.
  4. Turn-taking — timing logic decides when the agent should listen, speak, or yield, so the exchange feels natural rather than walkie-talkie.
You don’t wire these together yourself. You configure the agent’s behavior; the runtime orchestrates the pipeline.

​
What you configure today

The system prompt is the configurable surface in this release. Voice, LLM, tools, and knowledge base have placeholders in the dashboard but are not yet wired — they land in later releases. Document only what you can actually change so your team isn’t surprised.
  • System prompt — the agent’s persona, goal, tone, and guardrails. This is where almost all of an agent’s behavior comes from. See System prompt.
  • Branches & versions — every saved change is a per-user draft, every publish is an immutable version, and you can fork a branch off any version to experiment without touching what’s live. See Versioning.

​
A typical loop

1

Create an agent

Give it a name. It starts on its Main branch with a first version.
2

Write the system prompt

Edits autosave to your draft on the current branch — nothing ships until you publish.
3

Preview it

Open the live preview and actually talk to the agent. Iterate on the prompt until it behaves.
4

Publish or branch

Publish a version on Main when it’s ready, or fork a branch to try a riskier change in isolation and merge it back later.

​
Next