Connected agent

A Connected agent is built and runs on another platform: Dialogflow CES or ElevenLabs Agents. Its prompt, knowledge and conversation live there. Studio connects it to the phone by its id and keeps everything around the call: phone numbers, transfers, the secure detour, recording, the tools it asks for, and call history.

A Connected agent has no Prompt or Knowledge section in Studio. Its settings are in the agent’s Connected agent section; transfers, the detour and hangup are in Transfer & hangup; and the client-side tools it calls are matched in Tools.

Credentials

Platform	Credential	Where in the jambonz portal
Dialogflow CES	a Dialogflow CES credential holding the Google service account	LLM Services, Hosted agents
ElevenLabs Agents	an ElevenLabs Agents key	LLM Services, Speech-to-speech

Each is its own credential. A Google speech credential is never used for CES, and an ElevenLabs speech credential is not an ElevenLabs Agents credential. You choose the platform in the New agent dialog; if the account has no credential for it yet, you can still build the agent and add the credential later. Changing Platform later clears the credential and the agent.

Dialogflow CES

In Connected agent, choose the Credential and fill in:

Field	What it is
Project	The Google Cloud project.
Region	The agent’s region (`us` to start).
Agent (CES app id)	The app id from the CES console.
Deployment	Optional. Empty uses the latest.
Language	The conversation’s language (`en-US` to start).

Under Options:

Voice: Vendor voice (Google synthesizes the agent’s speech, the default) or jambonz voice (text comes back from CES and a text-to-speech credential in your account speaks it; choose the credential and voice).
Welcome event: sent when the call connects, with the variables as parameters. Default welcome.
If the caller says nothing for and send this event: after that many seconds of silence (up to 60), send the event (default no_input) to the agent. Off at 0, the default.
Caller can interrupt the agent: barge-in on the vendor’s audio. On by default.

The call ends when your CES agent ends its session; Studio hangs up and records the reason CES gives.

Client-side tools

In Tools, under Client-side tools your CES agent asks for, list the client-side tools your CES agent can call. Each name is matched to a tool on this Studio agent, or to the transfer or detour tool named in Transfer & hangup. A name with no match blocks publishing; Create tool starts a Studio tool with that name. Server-side tools run inside Google and are not listed.

In the CES agent’s instructions, have it call a tool rather than announce it, and give it no line to say after a transfer or detour tool: told to speak after a call, CES sometimes speaks and skips the call.

ElevenLabs Agents

In Connected agent, choose the Credential, then the Agent. Studio lists the agents that key can run in ElevenLabs.

When you choose an agent, Studio reads from ElevenLabs:

its name
its audio formats, for what it listens in and speaks in. They must be raw PCM (for example pcm_16000, which suits phone calls); any other format blocks publishing until you change it in ElevenLabs.
its client tools
whether it allows its first message to be overridden (see A returning caller)
its voice

Studio shows what it read under the agent. If you change the agent in ElevenLabs, select Read the agent again.

The agent’s prompt, voice, greeting and tools are set in ElevenLabs. Studio carries out the client tools the agent calls: those matched to tools on the Studio agent in Tools, plus the transfer, detour and hangup tools below. A client tool with no match blocks publishing.

The call ends when the agent ends the conversation (its end call tool) or calls hangup.

Client tools named as in Transfer & hangup

A Connected agent asks Studio for a transfer, the secure detour or a hangup by calling a client-side tool. Studio answers the tool and carries out the action with the settings in Transfer & hangup. The tool names in your CES or ElevenLabs agent must match the names there (the labels say CES or ElevenLabs in place of platform):

Action	Turned on with	Tool name	Default
Transfer	Transfer to a person	*Tool name in your platform* agent**	`transfer_to_human`
Secure detour	Secure detour and return	*Detour tool name in your platform* agent**	`secure_detour`
Hang up (ElevenLabs Agents only)	Let the agent hang up	fixed	`hangup`

Transfer: When the agent should transfer the call becomes the tool’s description. Under Add this tool to your platform agent, Studio shows the Client-side tool definition to add: its name, description and parameters. The parameters include the destination when the destination is Chosen by the agent, and a summary for the brief on a warm transfer.
Secure detour: a client-side tool with no parameters.
Hang up: a CES agent ends the call by ending its session, so it needs no hangup tool.

For an ElevenLabs agent, when one of these is turned on and the agent has no client tool with its name, Studio notes that the agent cannot transfer, take the detour or end the call. Add the client tool in ElevenLabs, then select Read the agent again.

Variables

A Connected agent has no prompt, so its variables are declared in Connected agent, under Variables go in as session parameters (CES) or Variables go in as dynamic variables (ElevenLabs). Each has a Name and comes From a Fixed value, a SIP header or the Dialer payload. Up to 20.

CES receives them with the welcome event, as session parameters.
ElevenLabs receives them when the conversation starts, as dynamic variables; use them as {{name}} in the agent’s prompt and first message.

caller_phone, dnis and call_sid are always sent.

Studio’s own lines

Studio speaks some lines itself, with a text-to-speech credential from your account: the brief on a warm transfer, and for a CES agent the detour message (Caller hears first). An ElevenLabs agent says the detour message itself.

ElevenLabs Agents: Voice for Studio’s own lines is below the agent. When the agent has a voice, Speak them in the agent’s voice speaks Studio’s lines in that voice, with an ElevenLabs text-to-speech credential. When you first choose an agent that has a voice and your account has an ElevenLabs text-to-speech credential, Studio turns this on for you.
Dialogflow CES with Vendor voice: choose a text-to-speech credential and voice under Voice for Studio’s own lines, in Options. With jambonz voice, Studio’s lines use that voice.

A returning caller

A caller can come back to the agent: from a transfer that does not go through, when If the transfer does not go through is set to Back to the agent, or at the end of the secure detour.

Dialogflow CES: the caller returns to the same CES session.
ElevenLabs Agents: there is no session to return to, so Studio connects the caller to the agent again. If the agent allows its first message to be overridden, Studio leaves the first message out, and the returning caller is not greeted again. If it does not, the caller hears the first message again; with a transfer or the detour turned on, Studio notes this. To fix it, allow the first message to be overridden in the agent’s Security settings in ElevenLabs, then select Read the agent again.

Realtime voice API: build the agent in Studio, with one vendor running the voice loop
Getting started: credentials and your first test call