Agents

An agent combines instructions, business knowledge, voice settings, and conversation rules. Use separate agents when teams have different goals, audiences, or escalation paths. This keeps conversations predictable and makes testing easier.

​

Create an agent

1. Open AI Voice Agents and select Create Agent.
2. Choose Create from Template for a predefined starting point, or Create from Scratch for a fully custom agent, then select Continue.
3. In Enter Agent Details, complete the required fields and select Create Agent.
4. Connect or create the knowledge base the agent should use from the Knowledge Base tab.
5. Complete the configuration panels under Agent Settings using Save & Next.

The agent list shows Agent | Language | Industry | Created On | Status (Active / Deactive) | Actions with Edit Agent and View Agent Details. Use search by name and Rows per page 10 / 15 / 25 / 50 for long lists. If Create Agent is disabled, you have reached your plan agent limit or have view-only access. Give the agent a name that identifies its purpose and audience. Add a description that helps team members understand where it is used and who owns future updates.

​

Enter Agent Details reference

A new agent automatically gets a knowledge base named <AgentName> KB and default conversation services. You can change them later from the configuration panels.

​

Choose the right settings

​

Complete the configuration panels

Open the agent to find top tabs Agent Settings | Knowledge Base | Test Agent, plus Add Integration and Conversation History. The left ConfigSidebar contains Agent Details | LLM Model | Conversation | Voice | Transcriber. Each panel has Save & Next and shows Incomplete Details until required fields are complete. Default selections for a new agent use the Valuez voice stack with balanced response settings. Ganas voices may show Setup required until an administrator configures the API key. Use Save & Next to move through the settings. Required fields must be complete before some tests and calling features become available.

​

Validation messages

If a panel is incomplete, the agent shows which section needs attention:

• Agent Details: language, region, time zone, industry, icon, or instructions missing.
• LLM Model: choose an AI provider and model.
• Conversation: greeting and ending messages required; transfer number must contain 7–15 digits when transfer is enabled.
• Voice: provider, voice, and speech rate required; custom rate must be 0.7–1.2.
• Transcriber: choose a speech-to-text provider and model.
• System Prompt: instructions required with at least 10 words.

​

Define conversation boundaries

Tell the agent which requests it can complete, which information it must collect, and when it should transfer or end a call. Include clear rules for unavailable information, sensitive topics, and repeated misunderstandings. Write boundaries as allow / collect / escalate lists. Example allow list: answer hours and pricing from knowledge, collect name and callback number, escalate billing disputes and angry callers to the human number. Anything not on the allow list gets the fallback response, never an invented answer. Review boundaries with the team that receives escalations so the transfer actually resolves the caller’s problem.

​

When to split into two agents

Create a second agent when any of these differ: audience language set, product line and knowledge sources, escalation destination, or success metric. A support agent measured on resolution and a sales agent measured on qualification should never share one prompt - their instructions, knowledge, and transfer rules pull in opposite directions. Linked cost: two agents double testing, so split only on a real difference, not a hypothetical one.

​

Plan the full conversation

1. Opening - identify the service and invite the caller to explain their need.
2. Understanding - ask only the questions required to select the correct path.
3. Resolution - answer, collect information, or perform the intended handoff.
4. Confirmation - repeat important details and confirm the next step.
5. Closing - end clearly so the caller knows the conversation is complete.

Define how the agent should respond when the caller is silent, interrupts, changes language, or gives incomplete details. Natural handling of these moments is as important as the main scripted path.

​

Keep the experience focused

• Use a short greeting.
• Give the agent one clear purpose.
• Define what it should do when it does not know an answer.
• Add a human transfer path when needed.

​

Duplicate, archive, and restore

There is no version history slider: your safety net is naming discipline plus prompt copies. Before a risky change, duplicate the working by copying the current instructions, greeting, and ending message into a dated note (for example Support-IN-2026-10-01). After a bad change, paste the saved wording back and retest one focused plus one end-to-end call. Rename retired agents with a ZZZ-Archived- prefix instead of deleting them, so old call records stay interpretable and no campaign accidentally selects a dead agent.

​

Manage changes safely

Before editing a live agent, note the current behavior and the calls affected by the change. Update the prompt, knowledge, or voice setting that directly addresses the issue, then test the complete caller journey. Review the first live conversations after a major change. If results are worse, restore the previous wording or setting while you investigate. After changing an agent, run a focused test for the updated behavior and one full end-to-end call. Review the transcript to confirm that the spoken response matches the intended instructions.