Use voice chat with an agent

Start a voice conversation

Open a deployed agent’s chat and choose Start voice call. Review the consent state, allow microphone access, and confirm the agent identity before the conversation begins.

What the agent must have

Voice agents use the autonomous-agent SDK configuration. The agent needs an enabled chat trigger, a signed loop with the tools it may use, and speechToSpeech settings on that chat trigger. The live speech model is separate from the loop model that reasons and chooses tools.

Choose voice governance settings

Deployment and session startup also check provider capabilities, approved voices, residency, and route policy. Only tools in the signed loop are available during the call. See Build and optimize voice agents for the complete SDK example and tuning guidance.

• providerId and model select the managed realtime provider and speech model; the current managed path is OpenAI gpt-realtime-2.1.
• voice selects the approved spoken voice. If you do not choose one, the SDK uses marin.
• requestedMode uses strict by default so audio and transcript wait for applicable DLP checks; native lowers latency but can release audio before final-response DLP completes.
• reasoningEffort tunes realtime reasoning from minimal through xhigh; the default is low.
• recording is disabled by default. optional still requires tenant and user permission.

Stay in control

Mute, end, or restart the call from the voice panel. Iron Gorilla keeps a transcript and can pause for approval when a tool action needs review.

If live voice is unavailable or a safety check needs a stricter path, the conversation can fall back to a turn-based voice or text experience.