Understand the agent editor

The editor changes a working draft rather than the configuration serving production traffic. See Edit and autosave agent drafts and Publish and manage agent versions for the complete lifecycle.

The Wirevox agent editor is where you prepare, test, and publish an agent's behavior. This guide explains the current editor for inbound voice agents.

note

AI Chat Agents are in Beta and use a different editor. Some concepts are shared, but the controls and deployment workflow are not identical.

The four layers to keep separate

The editor displays information from several lifecycle layers at once.

Layer	What it contains	How it changes
Agent identity	Name, avatar, agent ID, channel, status, and phone-number connection.	Updated by their individual controls.
Draft	Editable prompt, model, voice, variables, knowledge selection, and conversation configuration.	Saved automatically while you edit.
Published version	The reviewed configuration used for live operation.	Replaced only when a draft is published.
Runtime status	Whether the published agent is Active or Paused.	Changed with Start Agent or Stop Agent.

An autosaved draft is not automatically published. A published agent is not automatically active.

Editor layout

The inbound editor has three main regions:

1. Agent header — identity, connection, estimated cost, save state, and lifecycle actions.
2. Prompt Editor — the main system prompt, variable and function insertion, warnings, and custom greeting.
3. Right-side panels — Configuration, Functions, Infolist, and browser testing.

On smaller screens, the right-side panels appear as drawers over the editor.

Agent header

Identity and connection

The left side of the header displays:

• the agent avatar;
• the editable agent name;
• an Inbound channel badge;
• an Unpublished changes indicator when the draft differs from the published version;
• the assigned phone number or No number assigned;
• a shortened, copyable agent ID; and
• the estimated voice cost in credits per minute.

Select the avatar to upload a PNG, JPEG, or WebP image. Select the edit icon beside the name to rename the agent. Agent names are limited to 25 characters.

The agent ID is the stable identifier used by routes and APIs. Copy it when support, an integration, or an API operation needs an exact agent reference.

Cost estimate

For voice agents, the header shows the current draft's estimated cost in credits / min. Open the cost breakdown to see its platform, speech-to-text, model, and text-to-speech components.

When draft and published configurations have different costs, the breakdown can also show the currently published rate. This lets you review the billing impact of a model, voice, or speech configuration change before publishing it.

The displayed rate is a configuration estimate. Actual usage depends on measured conversation duration and any separately priced function use.

Save state

The header can display:

• Saving... while a draft request is in progress;
• Save failed when the latest draft could not be saved; or
• the time of the most recent successful save.

Configuration changes are normally autosaved after a short delay. Leaving the editor, opening a test, publishing, or changing runtime status attempts to flush pending changes first.

important

Do not treat text visible in the editor as saved until the save state confirms success. Do not treat a successful save as a publish.

Prompt Editor

The system prompt is the agent's primary authored instruction set. It should define identity, objective, conversation behavior, factual boundaries, function rules, escalation, and safety requirements.

Insert a variable

Select {{variable}} in the prompt toolbar to insert one of the variables defined under Configuration → Variables.

Variables use double braces:

Our business hours are {{business_hours}}.

The warning indicator reports a variable reference when no variable with the same name is configured.

Insert a function reference

Select [[Function: action]] to insert an attached function or integration action. Function references use this format:

[[Function: check_availability]]

The editor can also suggest available functions when you begin typing a function marker. The warning indicator reports references that are not attached or refer to a disconnected integration.

Attaching a function and referencing it are related but separate:

• attachment makes the capability available to the agent;
• the prompt tells the agent when, why, and under which conditions to use it.

Prompt warnings

The warning menu can identify:

• variable — the prompt references a variable that is not defined;
• function — the prompt references a function that is not attached; or
• disconnected — the referenced integration provider is not connected.

Use Add or Attach from the warning list where available, or remove a reference the agent should not use.

Warnings help find mismatches, but they do not prove the complete prompt is correct. Test the intended call flow and every consequential action.

Custom Greeting

Select Configure beside Custom Greeting (Optional) to provide a fixed opening that is sent when an inbound call connects. If no custom greeting is configured, the voice runtime allows the agent to produce its opening from the system prompt and call context.

Keep the greeting short. Put the complete behavioral instructions in the system prompt rather than trying to encode the entire call flow in the greeting.

Template setup

Some template-created agents show recommended capabilities above the prompt. This area can link to function creation and attachment, and it marks recommendations that have been configured.

Recommendations are informational. They do not automatically create a function, connect an integration, or make an external action safe to use.

Configuration panel

Select Configuration in the header to open agent settings. The panel begins with Pipeline Architecture, followed by collapsible sections.

Pipeline Architecture

Choose between Modular and Realtime. Switching architecture selects the stored model and voice configuration for that architecture. It can also change which controls are supported.

Review and retest the model, voice, language, conversation timing, and cost whenever you switch architecture.

Configuration sections

Section	What it controls
General	Timezone and the model available for the selected architecture.
Variables	Named values inserted into the prompt with {{name}}.
Call Settings	The delay before an inbound call is answered.
Language & Voice	Primary language and voice, optional secondary-language settings, and supported speech-to-text options.
Voice Delivery	Voice personality and expressiveness supported by the current provider.
Audio Environment	Optional ambient audio preset and presence level.
Knowledge Base	The workspace knowledge base selected for retrieval during calls.
Conversation Settings	Supported response timing, speaking speed, interruption handling, filler filtering, silence behavior, auto-hangup, and call-duration limits.

Controls are capability-aware. A setting can be hidden or limited when the selected architecture, model, voice, or provider cannot execute it.

General

Set the timezone used for time-sensitive behavior and choose the conversational model. In Realtime mode, changing providers can also select a compatible realtime voice.

Variables

Each custom variable has a name and value. Names accept letters, numbers, and underscores. The editor inserts the variable as {{name}}.

Variables are useful for facts that should be maintained separately from prompt instructions, such as hours or service areas. A variable is not a secret-management system; do not place credentials or private access tokens in prompt variables.

Language & Voice

Changing the primary language can replace an incompatible voice with a supported default. Secondary-language voice switching is available only when supported by the current runtime configuration.

Advanced speech-to-text options are also constrained by language and architecture. For example, an English-only provider cannot be selected for a multilingual configuration.

Knowledge Base

Select Add Knowledge Base to attach one existing workspace knowledge base to the agent configuration. The panel shows the selected base's name, source count, and size and allows it to be removed from the draft.

Knowledge attachment does not replace prompt instructions. The prompt should still explain when the agent should consult knowledge and how it should respond when no reliable result is returned.

Conversation Settings

Conversation controls affect timing and audio behavior, not the agent's business policy. Availability depends on runtime capabilities.

Particularly important controls include:

• Response Speed — how long supported speech recognition waits before treating the caller's turn as complete;
• Speaking Speed — the requested voice pace, within provider-specific limits;
• Interruption Sensitivity (Words) — how much caller speech is required to interrupt the agent when supported;
• Silence Prompt Timeout — when the agent checks whether a silent caller is still present;
• Auto-Hangup Timeout — how long pure silence is allowed before the call ends; and
• Limit Call Duration — an optional lower cap; when disabled, the interface reports a 40-minute system limit.

Change one behavior at a time and test with realistic speech, background noise, interruptions, and silence.

Functions panel

Select Functions in the header to view capabilities attached to the agent. The panel groups them by execution stage:

• Pre-Call;
• During Call; and
• Post-Call.

Select Add Function to open the workspace function catalog. You can filter by function category and attach an available function with Add to Agent.

Integration functions whose provider is not connected are marked Unavailable or Disconnected. Connect the required integration before relying on those actions.

Some function cards display a per-use credit price. That charge is separate from the voice rate shown in credits per minute.

Removing a function detaches it from the agent. Review the prompt afterward and remove or replace any instruction that still references the detached function.

warning

Function changes are part of the draft lifecycle. Browser tests can exercise the draft, but live operation continues using the published snapshot until you publish the updated agent.

Infolist

Select Infolist to inspect what information the current configuration collects and where it goes.

The panel can show:

• system-managed caller and call fields;
• fields required by attached integration actions;
• custom information collected by functions; and
• destinations such as Contacts, Notes, Webhook, SMS, Email, or Transfer.

Infolist is an inspection view. It is not where destinations or function behavior are configured. Change the applicable function or integration configuration, then reopen Infolist to review the result.

Test panel

Select Test to open a live browser voice test. Wirevox flushes pending draft changes before opening the panel.

The test uses the current draft and can show a live transcript. It requires microphone permission and usable workspace credits but does not require a published version or assigned number.

Connected tools are enabled for an owner test and can perform real external actions. Use safe test data and confirm external results after the test.

Publish, start, and stop

Publish

Select Publish to promote the saved draft. The publish dialog accepts an optional version name and note. Wirevox validates the draft and blocks publishing when it contains configuration errors.

After a successful publish, the button is disabled until another unpublished change exists.

Discard Draft

When a published agent also has unpublished changes, Discard Draft becomes available. It replaces the current draft with the live published version and permanently removes the unpublished edits.

Start Agent and Stop Agent

Start Agent activates the published configuration. An inbound voice agent cannot start without:

• a published version; and
• an assigned phone number.

Stop Agent changes an Active agent to Paused. It does not delete the draft, published versions, or assigned number.

More menu

The three-dot menu contains actions with different lifecycle effects.

Duplicate

Duplicate creates a new paused agent from the source agent's current draft when one exists. It copies the draft configuration, attached functions, tools, knowledge-base association, and avatar.

The duplicate receives a new agent ID, a generated copy name, no assigned phone number, and no published version. It must be reviewed, published, assigned, and started separately. Workspace agent limits still apply.

Version History

Version History lists published and archived versions, including their version number, publish time, optional name, and optional note. The current published version is marked Live.

Restoring an older version overwrites the current draft. It does not immediately change the live version. Any existing unpublished changes are lost, and the restored draft must be reviewed and published before it becomes live.

Delete Agent

Delete Agent permanently removes the agent and its associated data. The interface treats this action as irreversible. Confirm the exact agent before proceeding.

Recommended editing sequence

For a controlled change to an existing agent:

1. Review the currently published behavior and logs.
2. Change the smallest necessary part of the draft.
3. Wait for the save state to confirm success.
4. Resolve prompt warnings and configuration errors.
5. Test the draft, including connected functions.
6. Review the draft cost estimate.
7. Publish with a useful version name and note.
8. Confirm the agent remains in the intended Active or Paused state.
9. Monitor new conversations for the expected result.

Next steps

• Follow Create your first voice agent for the complete initial setup workflow.
• Learn how to write an effective system prompt, including model-specific guidance and complete function rules.
• Compare models and pipeline architectures before changing an agent's model or voice path.
• Configure voices, languages, and speech, including recognition, bilingual routing, delivery, and testing.
• Tune conversation behavior and call controls, including interruptions, silence, duration, greetings, and transfers.
• Connect the published agent through phone numbers and call routing.
• Configure variables and runtime context without duplicating caller, date, knowledge, or provider context in the prompt.
• Add and test functions and integrations with least-privilege actions and complete prompt rules.
• Build and ground agents with knowledge bases and RAG, including source inspection and retrieval evaluation.
• Review How Wirevox is organized for draft, published, and workspace boundaries.