Agents
An agent is a configured AI persona: a system prompt, a voice, a set of languages, optional tools, and concurrency caps. Each call is run by exactly one agent. Workspaces can have many agents (sales, support, after-hours...).
Anatomy
- Name & persona — e.g.
Ava — Customer Support. - System prompt — multi-line instructions that define behaviour. We provide 20+ industry templates at
/workspace/agentsas starting points. - Opening greeting & closing phrase — spoken verbatim at call start and end. Use the closing phrase to trigger farewell-detection (the engine hangs up after the agent says them).
- Language — primary language. Voxy AI auto-detects code-switched languages within the call.
- Voice — one of Voxy's curated voices (Aoede, Charon, Fenrir, Kore, Puck, ...).
- Tools — enabled tool callbacks (lookupContact, bookAppointment, noteCallReason, escalateToHuman, etc.).
- Max concurrent calls — hard cap on simultaneous calls for this agent. The dialplan + service-layer guard both enforce.
- Recording disclosure — auto-prepended to the first agent turn. Required by most jurisdictions including GDPR for EU callers.
Prompt template variables
System prompts support {{handlebars}}-style placeholders:
{{caller_number}}— E.164 of the inbound caller{{contact_name}}— if the caller matches an existing CRM contact{{today}}— YYYY-MM-DD on the server{{platform_name}}— the workspace name
Placeholders are HTML-escaped + length-capped to block prompt injection via inbound caller metadata.
Editing & testing
The agent editor at /workspace/agents includes a “Run test conversation” panel that pipes the system prompt through Voxy AI and renders the result side-by-side. Useful for iterating prompt without burning real call minutes.

