Overview
The Agent Configuration page is where you define everything about how an agent behaves: its prompts, voice, workflow, tools, and settings. Every change is auto-saved and nothing goes live until you publish.
Contents
Agent Configuration Navigation Panel
Voice, Language, LLM & Behavior
Agent Configuration Navigation Panel
The Navigation Panel options are grouped in the same way options are grouped on the homepage. These options open different pages for fully configuring your AI Agent including the conversation workflow, tolls, guardrails and knowledge.
Configure menu
The Configure section of the left panel organizes these settings into separate pages, from the agent’s core prompts and voice to the tasks it can handle, the documents and tools it can use, and the boundaries it must stay within. Every change is auto-saved.
| Section | What it's for |
| Agent | The agent's prompts, voice, and behavior. |
| Procedures | Teach the agent how to handle a specific task, written in natural language. |
| Workflow | The full conversation graph the agent follows. |
| Branches | Separate lines of the agent's configuration. Each branch has its own draft and version history, and live traffic can be split across branches by percentage. |
| Knowledge Base | Documents the agent can reference during live calls (using RAG). |
| Analysis | Configure how post-call analysis is performed, including evaluation criteria for successful conversations, custom data points to extract from transcripts, and optional sentiment analysis and topic discovery. |
| Tools | Webhooks, client actions, and integrations the agent can use during calls. |
| Guardrails | Boundaries that control what the agent can say and do. |
| Settings | Security, compliance, and advanced behavior. |
Monitor menu
The Monitor section of the navigation panel helps you see how your agent is performing once it is handling calls. Use these pages to check response speed, review past conversations, and run tests to make sure the agent behaves the way you expect before and after you publish changes.
| Section | What it's for |
| Observability | Average latency across each stage of a response: from the caller's last audio, through request dispatch, model generation, and speech synthesis, to the end-to-end response. |
| Conversations | Conversations handled by this agent. |
| Tests | Run tests from your library or create new ones to validate the agent's behavior. |
Deploy
The Deploy section provides a place to set which channels the Agents are deployed on. Click the + Request a channel button to automatically open an email form to request a channel from the Convoso Voice AI Agent team.
Click on an available channel to open it's configuration page and connect the channel to you Voice AI Agent system and add phone numbers to those channels.
Agent Configuration Page
When you open an agent, you’re taken to its configuration page, where you can build and fine-tune everything about how that agent works.
Control panel
The controls at the top right of the agent configuration page let you test, track, and release changes to your AI agent. Use them to manage variables, view and restore past versions, preview the agent with a test call, and publish your changes when they’re ready to go live.
| Control | What it does |
| Variables | Type {{ in the system prompt, first message, subagent prompts, or a Say node to create a variable (for example, the Lead's first name). Variables you create appear here. |
| Versioning | Changes are auto-saved as a draft. Publishing saves the draft as a new version. |
| Preview | Opens a call screen so you can place a test call and hear how the agent performs. |
| Publish | Publishes your changes. Click the down arrow for more options: publish to a branch, view all branches, view pending changes, or discard changes. The button is grayed out until there are new changes to publish. |
System Prompt & First Message
The system prompt defines the agent's personality, role, and instructions. Use it to describe the agent's behavior and tone and give it background and context.
A strong prompt usually covers:
Goal: What the agent is trying to achieve, such as qualifying Leads or booking appointments.
Environment: The business, campaign, audience, and situation the agent is working in.
Tone: How formal, friendly, or professional the agent should be.
Qualification criteria: Clear rules for what makes a Lead qualified, and the questions the agent should ask to confirm it.
First Message
The greeting when a call starts. Enter the opening statement an Agent makes when a caller answers the phone.
- Click the Interruptible toggle to make sure that a caller can interject when the Agent is talking.
Call Behavior
At the bottom of the Agent Configuration page there is a Call Behavior setting that can be used to add Background noise to your Agent's calls.
Activate this settings by selecting an option from the Background Sound drop-down menu. Click the preview button to hear the selection.
Options include:
- Office - Normal office noises; ringing phones, people talking.
- Crowded Room - many people talking in the background, no distinct conversation.
- City - Car horns and traffic sound effects.
- Forest - The dull sound of wind, water, and birds
- Keyboard - Computer keyboard typing.
- Hold Music - Relaxing instrumental music
- Custom upload - A sound file of your own.
Voice, Language, LLM & Behavior
Select the agent's voice, languages, and LLM from the panel on the right of the configuration page.
Agent Voice
The Agent voice panel controls how your agent sounds on calls. Open it by clicking the agent’s voice in the Voices section of the Agent page. From here, you can choose a voice and the engine and model behind it, add emotion and tone with expressive mode and audio tags, and fine-tune how consistent and natural the voice sounds during a conversation.
The settings available depend on the TTS model family you select, so you may not see every option below for every model.
Setting |
What it does |
|
|---|---|---|
A |
Engine | Selects the voice engine your agent uses, such as Voso 3.0. |
B |
Voice | Shows the voice your agent speaks with. Click the voice to browse and choose a different one, or click the play button to hear a preview. |
C |
TTS model family | Selects the text-to-speech model family, such as Emotive. The settings available below depend on the model you choose. For example, speed, style, and clarity boost don’t apply to the Emotive model. |
D |
Expressive mode | When turned on, lets your agent use audio tags to add emotion and tone to its speech. |
E |
Suggested audio tags |
When expressive mode is toggled on Audio Tags open. Tags that shape how your agent delivers a line, such as Calmly, Empathetically, Chuckles, or Sighs are shown. Click a tag to add it, or click Add to create your own. You can add up to 20 tags. |
F |
Stability | Controls how consistent the voice sounds throughout a call. Lower values sound more varied, but below 0.5 the voice’s accent and volume can change mid-call, so we don’t recommend going below that. |
G |
Voice similarity | Controls how closely the agent’s speech matches the original voice. Higher values sound closer to the original. |
H |
Natural numbers | When turned on, the agent speaks numbers the way a person would, such as reading dates, prices, and phone numbers naturally rather than digit by digit. (i.e., four-oh-five for 405) |
I |
Conversation reflexes | Sets how often the agent uses small, natural responses while the caller is speaking, such as “mm-hmm” or “okay.” Choose Off, Subtle, or Normal. |
Language
Select one or more languages that your Agent will be able to speak and understand fluently. Search for the language you want or scroll to find your desired language(s).
Click on a language and it will automatically be added to your list of languages, the panel will close after a language is clicked.
- A) Click on the top default language to change the default selection, the Language panel will open, click a new language and the default will instantly be changed, the panel will close again.
- B) Open the language panel to add a new language to the available list by clicking the + Add Additional language button.
C) Remove a selected language by clicking the X next to the name.
LLM
The LLM panel controls the large language model (LLM) that powers your agent’s responses. Open it by clicking the model in the LLM section of the Agent page. From here, choose the primary model and backup models, adjust how creative or predictable responses are, set how much the model reasons before answering, and manage limits and timeouts that keep calls responsive.
Setting |
What it does |
|
|---|---|---|
A |
Backup LLM configuration | Sets how backup LLMs are used if the primary LLM fails. Choose Default to use the primary settings, Custom to select multiple backups, or Disabled to turn off backups, disabling is strongly discouraged. |
B |
LLM | Shows the LLM your agent uses, such as Claude Sonnet 4.5, labeled Primary. Click it to choose a different model. |
C |
Temperature | Controls how creative or predictable the agent’s responses are. Move the slider left for more consistent, predictable responses, or right for more varied, expressive ones. Click Don’t send temperature to use the model’s own default instead. |
D |
Reasoning Effort | Sets how much the model reasons before responding. Choose Default, None, Low, Medium, High, or XHigh. Higher settings can improve responses to complex requests but may add latency. |
E |
Reasoning summary | When turned on, saves the model’s reasoning summary to the call transcript. This only applies to thinking-enabled models and may increase latency. |
F |
Limit token usage | Sets the maximum number of tokens the LLM can generate in a response. Enter -1 for no limit. |
G |
Soft timeout |
Lets the agent give the caller immediate feedback while it waits on a longer LLM response, so the call doesn’t go silent. Click to set how long the agent waits before responding and what it says. Example: “2s · LLM-generated” means a 2-second wait with a filler phrase generated by the LLM] |
Soft Timeout Settings
The Soft Timeout setting opens a panel with several toggles and configuration settings.
| Setting | What it does |
|---|---|
| Enable soft timeout | When turned on, the agent gives the caller immediate audio feedback if the LLM takes longer than expected to respond, so the call doesn’t go silent. |
| Timeout duration |
Sets how long, in seconds, the agent waits for the LLM before speaking a filler phrase, and how long it waits between fillers if the silence continues. The agent speaks as many fillers as fit before the recovery timeout, up to eight.
|
| Wait for first user message | When turned on, the agent doesn’t use fillers until the caller has spoken. This prevents fillers during the agent’s opening turn or a tool call before the caller says anything. |
| Use LLM-generated message | When turned on, an LLM writes the filler and recovery lines to match the agent’s persona and language. This is the same as the Speculative Turns setting under Settings > Advanced. Turn it off to write both sets of lines yourself. |
| View generated messages | Shows the filler lines generated for this agent. To review or regenerate them, go to Settings > Advanced > Speculative Turns. |
| Override soft timeout prompt | When turned on, replaces the default prompt used to generate filler messages with your own. Recent conversation context is still sent separately. Leave it off to use the default prompt. |
Failover & recovery
These Soft Timeout settings control how the agent switches to backup LLMs when a model fails or responds slowly. Failover applies to every channel, including voice, chat, widget, WhatsApp, and tests. Parallel racing and chain restarts apply to voice only. Leave a field empty to use the platform default.
| Setting | What it does |
|---|---|
| First token timeout |
Sets how long, in milliseconds, a model can go without responding before the agent fails over to the next backup. On text channels, this applies only if you set a value here.
|
| Failover patience |
Sets how long, in milliseconds, the agent waits for output before running the next backup in parallel. Whichever model answers first is used. Enter 0 to turn off parallel racing.
|
| Failed backup cooldown |
Sets how long, in milliseconds, a backup that just failed is moved to the end of the backup order instead of being tried first. Enter 0 to turn off the cooldown.
|
| Recovery timeout |
Sets how long, in seconds, a turn can stay silent before the agent stops every attempt and restarts the backup chain from the top. Fillers stop at this point, so this also limits how many are spoken. Enter -1 to turn off restarts.
|
| Recovery attempts |
Sets how many times the agent can restart the backup chain in one turn before it gives up and asks the caller to repeat themselves.
|
Agent Behavior
The Agent behavior panel controls your agent’s default response style and how it replies across text channels. To open it, select Agent behavior at the top of the panel. Changes are auto-saved.
Setting |
What it does |
|
|---|---|---|
A |
Default personality | When turned on, applies a standard response style to your agent. With this setting, the agent avoids repeating who it is unless asked, gives helpful and polite responses, asks clarifying questions when it needs more information, keeps answers clear and concise, and uses the caller’s first name sparingly and naturally (never at the start or end of a sentence). |
B |
Text channels |
Sets how your agent replies in each text channel. Click a channel to expand its settings.
|
Text Channel Settings
Setting |
What it does |
|---|---|
| Verbosity |
Sets how long the agent’s replies are.
|
| Output format | Sets whether replies are sent as Plain text or Markdown. Some channels only support plain text, so Markdown isn’t available for them. For example, WhatsApp delivers plain text only. |
| Response time |
Sets how quickly the agent replies. Realtime replies right away to keep latency low for live conversations. Choose 5 min, 10 min, or 1 hour to have the agent wait before replying.
|
Note: Once settings are changed a Reset to channel defaults option will show, click to reset all changes made back to the original Channel Settings default configuration for that channel.