Flow nodes
Use nodes to build the agent’s conversation flow. Each node represents a step in the dialogue or an action the agent can take.

- Conversation — lets the agent talk with the caller, ask questions, provide information, and continue the dialogue.
- Function — starts an external action, such as checking data or sending a request to another service.
- Code — runs a custom piece of logic when the standard nodes are not enough.
- Logic Split — sends the conversation in different directions depending on the caller’s answer or a condition.
- Extract Variables — finds important details in the conversation and saves them for later steps.
- Transfer — transfers the call to a human operator or another agent.
- Component — groups a reusable part of a flow so it can be used in several places.
- MCP Tool — connects the agent to an MCP tool that can perform a specific action.
- Override Settings — silently switches the voice, speed, language, volume, LLM or ASR for the rest of the call.
- End — finishes the conversation with a final message.
Begin node
Begin is the flow’s start node. There is always exactly one on the canvas, and it cannot be deleted. Its single connection leads to the first dialog node. The Begin inspector has three parts:- Who speaks first — two cards to choose from:
- AI speaks first — the agent says the first node’s reply as soon as the call connects. The usual choice for outbound calling: otherwise both sides stay silent.
- User speaks first — the agent stays silent and waits for the caller to speak.
- Start if the caller is silent — a switch with the seconds next to it, inside the User speaks first card; see below.
- First node — where the start connection leads. Change it here or drag the connection on the canvas; Show on canvas scrolls to that node. Without a start connection the inspector warns: “No start edge — pick a node here or drag a connection from Begin.”
Start after caller silence
In the User speaks first card turn on the Start if the caller is silent switch and set the caller’s silence in seconds next to it: 3 by default, 30 at most. A short timeline in the card shows what happens after the call connects. The Begin pill on the canvas shows the time you chose. If nobody speaks for that long after the call connects, the agent says the first node’s reply exactly as in AI speaks first mode: the node’s usual prompt,{{...}} variables are filled in. History shows the event Agent started after caller silence at that moment.
- The timer is cancelled if the caller speaks, if voicemail detection fires with the hang-up or leave-a-message action, or if the call ends. The timer fires at most once per call.
- What to set. 2.5–3 seconds leave time to detect a late voicemail or call screener before the agent’s first phrase. Voicemail detection keeps running during the greeting.
- Silence reminders in Call control do not change: they start counting after the greeting, and neither a reminder nor end-on-silence fires before it.
- After a transfer to another agent, the new agent’s Begin timer is armed again.
Conversation node
The Conversation node is used for a spoken interaction with the caller. It can ask questions, explain information, collect details, and send the conversation to the next node.
- Node name — in the inspector header; gives the step a clear name so it is easy to recognize in the flow. Delete node is in the ⋯ menu next to it.
- Reply — two tabs, Prompt and Static: generate the reply from an instruction, or speak a fixed phrase. Instruction (the LLM generates the reply) is what the agent should say or do at this step. Open in editor opens the reply in a full-screen editor; its Variables list inserts a variable such as
{{transfer_status}}at the cursor. - Skip response — a switch: speak and transition immediately; the node’s single transition then becomes Always.
- Global node · reachable from any node — a switch that makes the node reachable from any point in the flow.
- Transitions — define where the conversation goes next. An Always transition is unconditional and blocks other transitions from the same node. A Timeout (no reply within N seconds) transition fires when the caller stays silent for the set number of seconds; it is available only on the transitions of a Conversation node, at most one per node.
- Tools (the LLM calls them itself) — an optional list, added with + Tool. With tools, the model calls them on its own and a dialogue turn takes longer. + Transition tool adds a handover in the same turn — see below.
- Advanced — a collapsed block for the rarer settings; its title shows how many are set, and it opens by itself when something inside is set:
- Node-level overrides — a switch left off means the node inherits the agent defaults. Interruption, the generation model and the Voice — a different TTS voice while this node speaks — can each be overridden. The model and the voice are chosen from the catalogue only. The model parameters are in the gear, and the voice model and settings are in the voice window; see Node model and voice parameters.
- Knowledge bases — optionally select a knowledge base for this node instead of using the agent’s default knowledge base and search settings.
- Finetune examples — transcript → expected transition pairs that show the classifier where to take the conversation.
- Global node — expert options — on a global node only: the retrigger pause and Check on every client turn, see below.
Node model and voice parameters
A model chosen in a node’s overrides has a gear in its caption row — the same parameters window as the agent’s. It is there for the generation model, the argument-extraction model (Function) and the variable-extraction model (Extract Variables), the model of the transition decision (Logic Split), and the LLM, Classifier and ASR groups of the Override Settings node. A node voice is set up in the same voice window as the agent’s voice.- Model — the … — parameters window: Temperature, Answer limit, Service tier and, for a reasoning model, Reasoning effort — as for the agent’s execution model. A model that does not accept a temperature has no such field.
- Voice — the voice window, for the voice override of a Conversation node and for the Voice group of an Override Settings node. In it you pick the voice, the Model card if the voice allows several, and Fine-tune: for an ElevenLabs voice, Stability, Similarity, Style and Speaker boost (what they change), for a voice with a passport, the passport fields; under Developer, Parameters outside the passport (JSON). A value left empty is inherited — from the voice set by an Override Settings node, or else from the agent. Such a field is marked “inherited”; a set one has a button that resets it to inherited. Speaker boost has three options: inherited, On, Off. The node’s window has no Speed field: a Conversation node speaks at the agent’s speed or the one set by an Override Settings node, and an Override Settings node has a separate Speed group. The voice sample plays in the agent’s primary language. Picking another voice resets the node voice’s model and settings.
- Recognition — the Recognition parameters window, in the Override Settings node only: the ASR model parameters and Override provider-specific parameters (JSON).
Hand over to a specialist in the same turn
A transition along an edge is normally chosen at the start of the next turn, after the caller speaks. So once the agent decides a specialist is needed, it asks “Shall I connect you?” — and the caller silently waits to be connected instead of answering. A Transition fixes this: the model says its line and hands the call over in the same reply.- In the node’s tools section, click + Transition tool.
- Fill in Name and When to call — for example, “the question can only be solved by a specialist” — and pick the Target node.
- Only a Transfer node can be the target. If the flow has none yet, add one on the canvas first.
- A node without a target cannot be published, and nothing but a transfer node can be the target.
- Without a When to call description the editor warns you: the model has nothing to tell it when to call a specialist.
- If the transfer node has its own phrase before the transfer turned on, the editor warns you: the caller would hear two handover phrases in a row.
- The transition also works inside a component, with the same checks. It does not run in realtime mode.
A node with tools replies without streaming, so its first line may come a little later than usual. Check it with a test call.
Checking a global node on every caller turn
A global node is checked by the transition classifier. But on some turns the classifier is not called: when the node moves on by itself — through an Always transition, with Skip response ticked, or by an Equation condition. On those turns the global node is not checked, and a caller line can slip past it. This is how a global node failed to fire for a call-screening service that answers instead of a person and asks the caller to introduce themselves. The Check on every client turn checkbox on a global node checks it on those turns too. On a Conversation node it is under Advanced › Global node — expert options; on other nodes it sits under the Global node switch. The cost is one extra classifier call on each such turn. It is off by default, and the behavior stays as before.Function node
The Function node runs an external action, such as checking a customer’s data or sending information to another system. It can use a real HTTP request in production or a mock response while you develop and test the flow.
- Node name — identifies the function step in the flow.
- Name and Description in the Function block — what the function is called and what it does.
- Mock mode (deterministic response without HTTP) — a toggle between a predictable stub for development and a real HTTP call.
- Params (values support
{{example}}) — the function’s input values, with call variables substituted into them. - LLM arguments (parameters: name ← description) — arguments the model extracts from the dialogue and merges into the params. An empty field means
{{example}}substitution only. - Response variables (variable ← JSONPath) — map values from the function response into call variables. These variables can be used in later nodes and transitions, including success, error, and timeout branches.
- Speak during (reply while the call runs, optional) — what the agent says while the function is working.
- Mock response (JSON, supports
{{example}}) and scenarios — the status and body of the stubbed call, then add named success, error, or timeout scenarios. Mock scenarios keep tests from calling real endpoints. - Node-level overrides — a dedicated argument extraction model for this node. The override applies only to LLM-built arguments: plain
{{example}}substitution does not use the model, so there is nothing to override there. - Transitions — add a destination node for the function result. Use response variables to create success and error branches.
- Global node — make the function reachable from any point in the flow.
- Delete node — remove the function from the agent flow.
Code node
The Code node runs custom Python logic in a sandbox. Use it when the flow needs a calculation, data transformation, or another small action that is not covered by the standard nodes.
- Node name — identifies the code step in the flow.
- Code editor — write the Python code that the node should execute. The sandbox provides the
varsdictionary, theresultvariable, thenowclock and safe built-ins; there is no access to the network or the file system. What the code may not contain is listed under What code may not contain. - Current time —
now("Asia/Almaty")returns the time in an IANA time zone as a dictionary:epoch,iso,date,time(HH:MM),hour,minute,second,dow(1-7, Monday first),weekday(0-6, Monday first) andtz. The zone is required:now()without an argument, an empty string or an unknown zone raises an error and the node leaves through its error branch. Daylight-saving shifts are handled by the zone database, so a time-of-day check does not need a Function call to an external time service. - Output variables — list the names of values produced by the code, separated by commas. These variables become available to later nodes and transitions. Names that belong to the sandbox itself —
vars,result,now— are never exported: the validator warns about such a name, because the variable would always stay empty. - Execution timeout — sets the maximum time allowed for the code to run. The execution is stopped when the limit is exceeded.
- Transitions — connect the code result to the next node in the flow.
- Global node — make the code node reachable from any point in the flow.
- Delete node — remove the code step from the agent flow.
What code may not contain
The node’s code is checked before it ever runs — on publishing, on agent import and when a test starts. The validator names the node, the line and the construct: “node “Calculation”: the construct “import” is not allowed in flow code, line 3”.
Allowed built-ins:
abs, all, any, bool, dict, enumerate, float, int, len, list, max, min, range, round, set, sorted, str, sum, tuple, zip, the constants True, False, None, and the exception classes for try/except: Exception, ValueError, TypeError, KeyError, IndexError, AttributeError, NameError, ZeroDivisionError, RuntimeError, StopIteration, ArithmeticError, LookupError.
The same rules apply to code functions before and after the call, which additionally get metadata and http.
Logic Split node
The Logic Split node routes the flow silently according to equation-based conditions. It does not speak or wait for the caller; instead, it evaluates the available values and chooses the first matching transition.
- Node name — identifies the branching step in the flow.
- Silent branching — evaluates conditions without generating a reply or waiting for a new caller response.
- Transitions — define the possible destinations for each branch. A transition can use an Always, Equation, or Prompt condition.
- Evaluation order — conditions are checked from top to bottom. The order is: Always, Equation, Prompt, then Default. The first condition that matches wins.
- Always — routes the flow unconditionally and short-circuits the remaining conditions.
- Equation — evaluates a rule against known values, such as variables returned by previous nodes.
- Prompt — uses the model to evaluate a condition from the available conversation context.
- Default — acts as a fallback when no other condition matches. Add one to prevent the flow from getting stuck.
- Reorder transitions — drag transition cards to change their evaluation priority.
- Global node — make the split reachable from any point in the flow.
- Delete node — remove the branching step from the agent flow.
Extract Variables node
The Extract Variables node silently extracts structured information from the conversation. Use it when the agent needs to capture details such as a name, score, status, date, or another value for later steps.
- Node name — identifies the extraction step in the flow.
- Extracted variables — add one or more values that the node should find in the dialogue.
- Variable name — gives the extracted value a name, such as
var1. - Variable type — defines the expected format, such as
string. - Description for the LLM — explains what the variable means and helps the model identify the correct value in the conversation.
- Node-level overrides — optionally use a different extraction model for this node. If no override is enabled, the node inherits the agent’s extraction model.
- Transitions — connect the extraction step to the next node in the flow.
- Global node — make the extraction node reachable from any point in the flow.
- Delete node — remove the extraction step from the agent flow.
Transfer node
The Transfer node moves the call to a human operator or another agent. It supports a direct cold transfer and a warm transfer, where the agent can brief the recipient before connecting the caller.
- Node name — identifies the transfer step in the flow.
- Cold (blind) / Warm (with briefing) — connect straight through, or brief the recipient first. On a warm transfer the caller is put on hold, the agent dials the destination, hands over the summary and leaves. The summary is private — it can carry the caller’s name, amounts and terms — so the agent speaks it only once the caller is confirmed to be cut off from the agent’s audio. If that cannot be confirmed, the operator is connected without the briefing, and the transfer event in the call details says operator briefing skipped.
- Who hears the briefing — on a warm transfer, Operator only (the default) or Operator and caller. With Operator and caller the caller is connected to the operator as soon as the operator answers, and both hear the briefing; it cannot be interrupted and is recorded. Use it when the briefing is something the caller may hear, for example “I’m connecting Maria about her appointment tomorrow”. In this mode the call duration and the recording include the dialing time. If the briefing could not be generated, it is skipped and the call card says so.
- Destination — select where to send the call, such as a phone number, agent, or URI. A transfer to another agent works in every channel: a phone call, a call from your website, the voice test in the cabinet, chat and WhatsApp.
- Transfer variables — use values such as
{{transfer_status}}and{{transfer_success}}to track the result of the transfer. - Signaling method — choose SIP INVITE or SIP REFER for the transfer signaling.
- Transfer ring duration — set how long the platform waits for the destination to answer. The allowed range is 1-120 seconds.
- Transfer number — which number the recipient sees: Client’s number (default), Trunk number or Custom number. A custom number has a field of its own — a number or a variable; the value must belong to the outgoing number pool of the transfer subgroup, otherwise the transfer takes the fallback edge. With SIP REFER the number goes to the SBC as a hint in the
P-Asserted-IdentityandReferred-Byheaders — whether the carrier applies it to what the recipient sees is the carrier’s decision. - SIP headers — add custom
X-*headers. Header values support variables and environment secrets. - Transfer subgroup — picks the trunk line the transfer leg or REFER goes out on: its caller pool, limits, allowed destinations and REFER form. “Default” means the previous behaviour — the call’s own subgroup, otherwise the trunk default. If the five-digit prefix you pick is not provisioned on the call’s trunk, the dial falls back to the default subgroup and the reason shows up in History. The subgroups themselves are created in Telephony.
- Phrase before the transfer — the Speak a phrase before the transfer toggle: the node says “Connecting you to an operator…” itself, with no separate conversation node in front of it. The text is either fixed, with variable substitution, or generated from your instruction. The transfer only starts once the phrase has played in full; outside the operator schedule it is not spoken. The Do not let the caller interrupt this phrase checkbox covers every reply of the node, not just this one.
- Operator working hours — the Transfer only during working hours toggle limits handovers to a schedule. Set the Operators’ time zone — pre-filled with the organization time zone, change it if the operators sit elsewhere — and at least one row of days and hours; an enabled schedule with no rows cannot be published, and an interval that crosses midnight is written as two rows. Do not transfer in the last 5 minutes before closing keeps the node from handing over a caller nobody will pick up.
- What to say outside working hours — a fixed text or an instruction for a generated reply. An empty field means the node says nothing. The reply can use
{{transfer_next_open_time}}and{{transfer_next_open_date}}. - Call recording — the Keep recording after the transfer checkbox, off by default. By default the recording ends where the call duration ends — at the end of the conversation with the agent: the hold for a warm transfer, the start of the dial for a blind one. Turn it on to also record the caller talking to the operator: the recording then outlasts the call — mind your obligations around call recording. After a failed transfer the agent’s continued conversation with the caller is recorded and appended to the same file — see the recording in the call details.
- Failure variables — after a failed transfer
{{transfer_status}}holds the reason —no_answer,busy,rejected,transfer_ring_timeoutorerror— and{{transfer_success}}isfalse. Alongside them,{{transfer_failure_reason}}holds a technical code of why the transfer failed. Use them in fallback conditions. - Transitions — add fallback destinations for failed transfers. These transitions do not run after a successful transfer because a successful transfer is terminal for the AI.
- Global node — make the transfer reachable from any point in the flow.
- Delete node — remove the transfer step from the agent flow.
{{transfer_status}} becomes out_of_hours, {{transfer_success}} is false, the next opening is written to {{transfer_next_open_time}}, {{transfer_next_open_date}}, {{transfer_next_open_iso}} and {{transfer_next_open_epoch}}, the reply is spoken if you configured one, and the node leaves through its transitions. Give such a node a Default transition or a condition on {{transfer_status}} == "out_of_hours" — otherwise the validator warns that a call arriving outside the window has nowhere to go.
In the call history the transfer event is marked as skipped with the reason Outside working hours, so a closed window is not counted among the transfers that failed to reach an operator.
A transfer to another agent switches the voice too. When the destination is an agent, the continuation runs on the target agent’s settings: its voice and synthesis model, its languages, its silence reminders and backchannel words, and its call control limits. Silence timers start again from the first reply of the new agent.
Component node
The Component node inserts a reusable block into the agent flow. Use it for a repeated part of a conversation, such as a verification step, closing sequence, or common topic. When the agent is published, the component expands into its regular nodes.
- Node name — identifies the reusable block in the flow.
- Component — select an existing component to insert into the flow.
- Create component — when the block you need is not in the library yet.
- Transitions — connect the component to the next node in the surrounding flow.
- Global node — make the component reachable from any point in the flow.
- Delete node — remove the component reference from the agent flow.
Component prompt
The component editor has a Component prompt field — shared rules that every reply inside the component follows. Put there what you would otherwise copy into every node: for example, “do not name the debt amount until the client has confirmed their identity”.- The text is added to every agent reply inside the component, right after the agent’s global prompt.
{{var}}variables are substituted on each turn. - The component prompt does not define transitions: the transition classifier, variable extraction and function arguments do not get it. Transition conditions are still written in node instructions and on edges.
- The End node farewell, transfer lines, silence reminders and the voicemail message do not get it.
- The prompt is stored in the component: it is versioned with it, goes into a local copy and is shown as a separate section when comparing with the library. Component nodes without a prompt work as usual.
- The recommended length is up to 3,000 characters: the text is sent to the model with every reply inside the component.
MCP Tool node
The MCP Tool node calls a specific tool through an MCP server without generating a spoken response. Use it when the agent needs to retrieve information or perform an external action through an MCP integration.- Node name — identifies the MCP step in the flow.
- MCP server — the server that provides the tool. If no server is available, add one in the MCP servers settings. A server inside an internal network (
10.x.x.x,172.16.x.x–172.31.x.x,192.168.x.x, names ending in.localor.internal) is unreachable until its address is allowed for your organization — contact support. - Tool arguments — add the values that the selected tool needs. Arguments can include variables such as
{{var}}. - Response → variables (JSONPath) — map values from the tool response into call variables. These variables are available to later nodes and transitions.
- Tool result variables — use status, success, error, result, and HTTP status values to branch the flow after the call.
- Speak during — what the agent says while the tool is working.
- Timeout — set the maximum time allowed for the MCP call to complete.
- Require confirmation before the call — the agent asks permission first.
- Transitions — connect the tool result to the next node. Response variables can be used for success and error branches.
- Global node — make the MCP tool reachable from any point in the flow.
- Delete node — remove the MCP tool call from the agent flow.
Override Settings node
The Override Settings node changes the call’s settings for the rest of the call. It says nothing: when the call reaches it, the node switches what you enabled and moves on at once.
- After a language fork. The caller chose Spanish: the node switches the voice and the call language, and every later node — components, global nodes, the farewell, the transfer phrase — speaks Spanish with the new voice. Without it, every node with no voice of its own goes back to the agent’s default voice.
- “Please speak slower”. A global node with Apply and return lowers the speed and goes straight back to the node the conversation was on, which answers at the new pace.
- A model that handles the chosen language better, for the replies, the transitions or the speech recognition.
- Voice — the TTS voice for the rest of the call. The voice model, if there are several, and its settings are set in the voice window — see Node model and voice parameters. Changing the voice resets the model choice and the settings. A model that is not in the voice’s list gives a warning, and the node speaks with the record’s own model.
- Speed — Set a value, or Step from the current value, for example 0.1 slower each time. Min and Max keep repeated steps within bounds.
- Call language — one of the agent languages for the rest of the call. The voice, the recognition and the built-in phrases follow it.
- Volume — a gain in dB over the synthesized audio, set or stepped like the speed, within ±12 dB. It works for every provider.
- LLM — the generation and extraction model, for nodes without an override of their own.
- Classifier — the transition-decision model, for nodes without an override of their own.
- ASR — the speech recognition module. It is swapped in the background, and what the caller says during the swap can be lost, so place the node where the agent is speaking. A module that streams differently from the agent’s is refused in the call, and the agent’s recognition stays; the editor warns about this in advance.
Adding the node
- In the editor palette, pick Override Settings. The node is also available in the component editor: a switch made inside a component stays in effect after the call leaves it.
- In the node panel, in the Call settings override section, turn on the groups you need and fill in their fields. The voice and the models are chosen from the catalogue only; there is no manual entry. The model parameters are in the gear next to the chosen entry, and the voice model and settings are in the voice window. The node card on the canvas lists the enabled groups.
- Choose how the call reaches the node:
- By a transition. Put the node on the right branch and draw one transition out of it.
- As a global node. Turn on Global node, describe the Entry condition, for example “The caller asks to speak slower”, and tick Apply and return. No outgoing transition is needed in this case.
- Check the node in the agent’s validation tab; the findings are listed in the table below.
- Publish the agent. A voice, model or recognition module picked from the catalogs is pinned in the agent version the same way as the agent’s own. Bulk replace of a voice or model finds this node too.
Which setting wins
A node’s own override comes first, then Override Settings, then the agent’s settings. Each setting is compared on its own:
Several Override Settings nodes along the path add up: each one changes only its own enabled groups. If the first one set the voice and the second only the speed, the voice stays from the first one.
The switch does not undo itself. It lasts until the end of the call. To go back to the previous model or voice, add a second Override Settings node with the agent’s values further on. Only a transfer to another agent resets it: that agent starts with its own settings, and what Override Settings switched does not carry over.
Apply and return
On a global node, the Apply and return checkbox switches the settings and returns to the interrupted node at once, with no classifier decision. The interrupted node answers the same caller line, already with the new voice, pace or volume. It does not repeat its static phrase; it answers in its spirit instead. This works only on a global node. Retrigger pause. A global node has the Prevent immediate retrigger (suppression turns) field, 3 by default. Until the pause is over, the node does not fire again, so a second “slower, please” is skipped. For such requests set it to 1, the minimum. Even at 1, a request on the very next turn is skipped; one turn later it works.Limitations
- The call language does not translate texts. The node texts in the right language are still written by you or the LLM: the node changes the language of speech, not the content. The built-in phrases — silence reminders, wait phrases, active listening — switch to the new language; the voicemail message is said in the agent’s primary language.
- Only an agent language can be chosen. To switch the call to Spanish, add Spanish to Agent languages.
- Auto-switch to the caller’s language keeps working. If the agent has Auto-switch to the caller’s language on and the caller goes on speaking another language from the agent’s list, the call switches to that language. To keep the node’s language until the end of the call, turn the switch off.
- In chat and text channels — chat, WhatsApp, SMS, a text test session — only the LLM and Classifier groups apply. Voice, speed, language, volume and recognition have no effect there.
- In realtime mode the node does not run, and the editor warns about it. It works in cascade and flex modes.
- The agent’s fallback voice does not change. If the main synthesis is unavailable, the call goes on in the agent’s fallback voice, but the volume from Override Settings is kept.
- Volume is a gain on the finished audio, not a provider setting. It applies to every voice, including cached phrases and the fallback voice, and it is audible in the call recording. It does not change background audio. Do not confuse it with the Loudness gate in the agent’s voice settings: that one is about the caller’s audio.
Recipes
Language fork: the rest of the call in another language. After the node where the caller picks a language, split the branches with Logic Split. On the “Spanish” branch add Override Settings with the Voice (a Spanish voice) and Call language (es — Spanish must be one of the agent languages) groups, plus ASR if needed. Then lead the branch into the shared nodes and components: there is no need to duplicate them per language. Nodes before Override Settings speak with the agent’s voice.
“Please speak slower”. A global node with the Speed group: Step from the current value, value −0.1, minimum 0.8. Entry condition “The caller asks to speak slower”, Apply and return ticked, retrigger pause 1. Repeated requests never take the speed below 0.8.
“Please speak louder”. The same, with the Volume group: step +3 dB, maximum +6 dB. Do not go above +6 dB for now: the headroom before distortion on the phone line has not been measured yet.
A different model for a hard part. Before the objection-handling block, add Override Settings with the LLM group and a stronger model. After the block, add a second node with the agent’s model, otherwise the stronger model stays until the end of the call.
A different recognition model. Once the caller has chosen Kazakh, add Override Settings with the ASR group and a Kazakh-Russian model. A node where the agent speaks should follow right after it: the module switches while the reply is playing. The new model must deliver results the same way as the agent’s — streaming or in one piece — otherwise the swap is refused.
How to check
Voice, speed and volume can be heard only in a test call. A text test session shows only the LLM and classifier switch. In the call history, each run of the node shows up as a Call settings override event: it lists the enabled groups, the voice, speed, language, volume in dB and the models. A green event means everything applied. A yellow one means one group did not apply, and the row says why:- “TTS unit not built — the previous voice is kept”;
- “ASR not swapped: a different streaming kind — the agent ASR is kept”;
- “ASR module not built — the agent ASR is kept”.
Validation findings
Errors block publishing, warnings do not.End node
The End node finishes the conversation. Use it when the agent has completed the task and should say a final goodbye before ending the call.
- Node name — identifies the final step in the flow.
- Prompt / Static — generate the farewell from an instruction, or speak a fixed phrase.
- Fixed phrase — write the message spoken when the conversation ends. You can include variables such as
{{transfer_status}}or{{transfer_success}}. - Allow interrupting the farewell phrase — lets the caller interrupt the final message. When disabled, the phrase is spoken to completion and then the call ends.
- Delete node — remove the end step from the agent flow.
- Greeting — the first thing the caller hears
- Language and voice — set the default speaking language and voice style
- Prompt and flow — define how the agent should behave and what it should ask next
- Actions — transfer, end call, webhook events, and follow-up logic