A voice agent that attaches to a live call and speaks with the caller.

Attaches an agent to a live phone call and speaks with the caller. Audio streams directly to a real-time speech model, which allows the caller to interrupt the agent mid-speech. Configure the model, persona, prompt, tools, and context grounding indexes directly on the node.

Both call directions use the same three-step shape:

The node is labeled Voice agent in the Agent category of the Add node panel.

You may optionally configure structured output variables in the Outputs section of the properties panel. They are useful for setting goal-oriented or routing variables to use in downstream nodes within the workflow.

Each output variable you declare becomes a field on the agent's output object, referenced as $vars.<agentName>.output.<name>. Give each one a name and a type, then instruct the agent in the variable's description field how to populate it. Output generation is driven by a combination of the agent's system prompt and the output variable descriptions.

The agent populates them when it ends its voice session.

The two most common uses are call summaries and handoffs. For a summary, declare an output such as call_summary and use its description to tell the agent to summarize the call. The agent writes it when it ends its session, and downstream nodes read it as $vars.voiceAgent1.output.call_summary.

Connect tools to the Tools handle to give the agent capabilities it can call while reasoning. The agent decides during its loop whether and when to call each connected tool. Each tool carries a description that tells the agent what the tool is for, so write clear descriptions when a tool exposes one.

While designing tools for voice agents:

Every voice agent has access to two built-in tools. These are included but not shown in the canvas:

Tool-level guardrails are available on the tools connected to the Voice agent node. You can filter or block on a tool's input parameters. Configure the guardrail on the tool attached to the Voice agent's Tools handle. tool guardrails has more details.

Ending a voice session is not the same as ending the call. When an agent ends its session it exits the call but leaves the line open: the agent populates its output variables, and the workflow moves to the next node. Only the End call node and the caller can hang up the call.

That is what makes handoffs possible. Chain Voice agent nodes and bind them all to the same Call context. Each agent ends its own session when it is done, and only the last path reaches End call, so the caller stays on the line throughout.

A handoff has two parts:

The Voice agent has no Escalations connection, because a call can't wait while a person completes a task. To have a person act on the outcome of a call, add a node from the Human category, such as Quick Form, after the Voice agent.

To read a transcript after a call, open the execution trace of a test run, or the trace of a deployed run, as Monitor runs describes. Observability has more details.

To use the transcript inside the workflow:

For example, you can pass the transcript to an autonomous agent for processing:

The agent talks normally but never runs a tool. No unattended or default robot with a serverless machine is assigned where the workflow runs, so every tool call comes back as an error and the agent stays conversational. Ask your administrator to assign a robot, then call again. Folder requirements has more details.

The call ends before the conversation is finished, but there are no errors. The agent ended its own session mid-conversation, and the workflow moved on to End call. That is likely a prompting issue where the voice agent thought it finished helping the caller and decided to end the session.

The agent stalls or produces static reading digit strings. With Gemini Live at Temperature 0, long digit strings can stall or sound distorted. Raise Temperature slightly. If the issue persists, try the other model.

The agent fails to attach with a tool schema error. The Gemini model accepts a narrow schema subset. Flatten the tool's input schema, or wrap the tool in an API workflow with a flat input schema and use the workflow as the tool instead.

The call connects, nobody speaks, and it drops after about 25 seconds. A call that never attaches an agent is ended after about 25 seconds of dead air. Bind Call context to the callContext from the Incoming Call trigger or the Create Outgoing Call node.