Build your first voice agent

The fastest way to a useful first agent is to give Menace Voice a narrow brief and let the guided setup build the initial workflow around it. You do not need a phone number for this tutorial. You need a signed-in workspace, permission to use your microphone, and a working model configuration by the time you reach the Voice step.

If you have not configured models yet, start with Models and costs — set up Speech to Speech, Menace Voice-managed inference, or BYOK before the Voice step.

1. Start with a testable brief#

Job: help new and returning customers request a haircut appointmentSuccess: collect service, preferred day, preferred time, name, and callback numberBoundaries: never claim a slot is booked until the scheduling tool confirms itEscalate: transfer or capture a callback request when the caller needs staff help

A useful brief describes observable behavior. Avoid broad directions such as “be a helpful receptionist.” State what the agent should collect, what system result counts as success, and what it should do when it cannot complete the job.

2. Complete the six setup steps#

• Purpose: choose a starting template or Custom, name the agent, select inbound or outbound, and describe the job in detail.
• Knowledge: select existing documents or upload the policies and reference material this agent should use.
• Behavior: choose the tone and language, then add requirements, boundaries, escalation rules, and any wording the agent must preserve.
• Voice: confirm the organization model configuration and select the voice used by the generated agent. Self-hosted workspaces must supply valid managed or BYOK configuration.
• Connections: add only the built-in, HTTP, or remote MCP tools needed for this job. Templates can recommend appropriate options, but you control what is attached.
• Launch: optionally configure a pre-call caller lookup and a post-call webhook, review the summary, and create the first workflow.

What generation gives you

Menace Voice creates a workflow with an opening, ordered conversation stages, explicit transition conditions, and a closing path based on your brief. Connected documents and tools are attached to the generated conversational nodes. Treat this as a first draft to review, not a finished production script.

3. Run the first Web Call#

Open the generated agent and start a Web Call. Allow microphone access, speak as the intended caller, and watch the live transcript, node transitions, and tool activity. Web Calls run the speech-to-text, model, text-to-speech, workflow, recording, and transcript path without using a telephony provider.

• Happy path: provide the expected information and confirm the agent reaches the intended outcome.
• Missing information: omit one required detail and confirm the agent asks for it instead of guessing.
• Boundary: ask for something the agent is not allowed to promise and confirm it explains the limit or escalates.
• Interruption: speak while the agent is talking and check that the turn still feels natural for your interruption settings.
• Tool failure: when a tool is attached, exercise an empty result or failure and confirm the agent does not invent success.

4. Make one change at a time#

Edit the node that owned the incorrect behavior. Put shared persona, language, and non-negotiable rules in the Global node; keep stage-specific questions and tool instructions in the Agent node where they apply. Save the draft, start a new Web Call, and compare the new run with the previous transcript.

Keep the first version small

One clear job makes the transcript easy to evaluate. Add a new branch only when a real test shows that the current node cannot express the behavior cleanly.

When the conversation needs live data or an action, continue to Give your agent tools. For policies and caller-specific data, use Knowledge & context next.