Voice Agents

Voice Agents are the conversational AI workers at the heart of your Sysevo voice platform. Each agent is a self-contained "phone person" that answers and makes calls, talks to your customers in natural speech, follows the instructions you give it, looks things up, books appointments, and remembers who it spoke to. This guide walks you through everything you can do on the Your Agents page and inside the agent editor — from creating your first agent to fine-tuning its voice and embedding it on your website.

In this guide

• What a voice agent actually is and does, end to end
• The Your Agents page: the list, stats, status, and row actions
• Creating an agent (Agent Builder, Blank Canvas, and Import)
• The agent editor: Workflow view vs Simple view, saving and publishing
• Agent settings: name, language, voice, model, system prompt and first message
• Voice tuning controls (stability, speed, interruptions, silence)
• Caller memory and what your agent can remember
• Tools, Knowledge, Behaviour, Variables, Phone and Conversations sections
• The embeddable web call widget
• Testing your agent with a live web call
• Common workflows, tips, and troubleshooting

1. What is a voice agent?

A voice agent is an AI that can hold a real spoken conversation on a phone line or through a web widget. When someone calls (or the agent calls out), the agent:

1. Listens — speech is converted to text by a transcriber (STT, "speech to text").
2. Thinks — a language model (LLM) reads what was said, follows your instructions, and decides what to say or do next.
3. Acts — if needed, it uses a tool (transfer the call, look something up over a webhook, hang up, etc.).
4. Speaks — its reply is turned back into a natural human voice by a voice engine (TTS, "text to speech").
5. Remembers — if memory is enabled, it recognises repeat callers and recalls their details next time.

Every agent is built from the same building blocks, which you configure in the editor:

Note on terminology: Internally an agent is sometimes called a workflow. On screen you'll mostly see the friendly word Agent, but a few menus and messages (e.g. "Duplicate Workflow", "Import Workflow") still say workflow. They mean the same thing.

2. The Your Agents page

This is your home base. Open it from the Voice section of your portal sidebar (the URL ends in /voice/agents). The page title reads Your Agents with the subtitle "Deploy and manage conversational AI agents."

The page has four parts: a header with actions, a stats strip, an Active Agents list, and (if you have any) an Archived list.

2.1 Header and actions

If you have reached your plan's agent limit, the New Agent button changes: it shows a small padlock icon, and clicking it opens an Agent limit reached dialog instead of the create flow. That dialog tells you which plan you're on and how many agents it allows, with an Upgrade plan button. (More on limits in section 2.5.)

2.2 The stats strip

Just under the header is a row of live counts:

2.3 The agent list

Agents are shown in a clean table grouped into Active Agents and Archived sections (the archived section only appears when you have archived agents). Each row shows:

The little tag under the name reads live or archived, and shows · inbound or · outbound if the agent's name contains those words (this is just a friendly hint based on naming).

Clicking any row opens that agent in the editor. Hovering a row reveals quick actions on the right.

2.4 Row actions: edit, call, duplicate, archive, delete

Hover a row (or open the ⋯ menu on the right) to reach:

About status. An agent is either active or archived:

• Active agents appear in the Active list, count toward your Active Agents stat, and can receive calls.
• Archived agents are tucked away in the Archived list (shown at reduced opacity). Archiving is reversible — use it to retire an agent without losing its configuration or history.

Deleting is permanent and cannot be undone. The confirmation dialog reads "<Agent name> will be permanently deleted. This cannot be undone." If an agent is still used by a campaign, deletion is blocked with a message telling you to remove it from those campaigns first.

2.5 Plan limits

Your plan allows a certain number of agents. When you're approaching or at the limit, a plan limit banner appears near the top of the page. Once you hit the limit, the New Agent button locks (padlock icon) and clicking it shows the Agent limit reached dialog with an Upgrade plan link. Archiving an agent does not free up a slot for a new one in the way deleting does — deleting an agent removes it from your count.

2.6 If the list won't load

The Your Agents page talks to the Voice AI service in the background. If that service is briefly unavailable you'll see a soft amber banner: "Voice AI service is temporarily unavailable due to maintenance." The page automatically retries every 30 seconds, and there's a Retry button if you'd rather not wait. Your agents are safe — this is only a temporary connection issue.

3. Creating an agent

Click New Agent to open the full-screen create flow. The first question is "How would you like to start?" with the reassuring note "Both options create a draft — nothing is published until you deploy." You have two paths.

3.1 Agent Builder (recommended for most businesses)

Agent Builder is an AI-powered guided wizard. Instead of writing prompts yourself, you answer questions about your business and the platform builds a complete, industry-tailored agent for you. The wizard moves through these steps:

When you finish, the platform generates a full multi-step conversation flow — greeting, qualification, intent routing to specialist sub-flows, objection handling, closing, and call ending — pre-written for your industry. You can then open it in the editor and adjust anything.

Tip: Agent Builder gives you the fastest path to a genuinely useful agent. Even if you plan to customise heavily, it's usually quicker to start here and tweak than to start from blank.

3.2 Blank Canvas

Choosing Blank Canvas asks a second question — "Choose your default view" — with the note "You can switch between views at any time inside the agent editor."

Either choice creates a new agent pre-loaded with a sensible starter flow — a global node (overall goal and rules), a start/greeting node, a main agent node, and an end-call node — and drops you straight into the editor in the view you picked. The agent is given an automatic name like Voice-Agent-A1B2C3 which you can rename at any time.

3.3 Importing an agent

If you (or your provider) exported an agent as a JSON file, click Import on the Your Agents page. The Import Workflow dialog lets you drag-and-drop a .json file or click to browse. The file must be a valid agent export containing the conversation definition (nodes and edges). On success the agent is created and opened in the editor. If the file is invalid you'll see an error such as "Invalid workflow JSON — missing nodes/edges."

You can create an export to import elsewhere from inside the editor — see Download in section 4.2.

4. The agent editor

Opening an agent (by clicking its row, or Edit) takes you to the full-screen editor. The editor has a slim top header and a large working area. It supports two views of the same agent, which you switch between with a toggle in the header.

4.1 Workflow view vs Simple view

The header has a pill toggle with two options: Workflow and Simple. They edit the same agent — pick whichever is comfortable. Your choice is remembered per agent.

• Simple view — a clean, form-based interface organised into clearly labelled sections (Agent, Tools, Knowledge, Behavior, Variables, Phone, Conversations, Widget). This is the recommended view for most users and is the focus of this guide. Everything you typically need to configure is here.
• Workflow view — a visual canvas where each step of the conversation is a "node" and the lines between them are the paths the conversation can take. This gives you fine-grained control over complex, branching call flows. It's powerful but more advanced; the Simple view writes to the same underlying flow.

You can switch freely at any time without losing work.

4.2 Saving, publishing and other header actions

Changes you make in the editor are not live until you save and publish. The header gives you:

The save → publish cycle:

1. Make your edits in Simple or Workflow view.
2. Click Save — your work is stored as a draft (you'll see "Workflow saved as draft"). The unsaved-changes badge clears.
3. Click Publish — the draft becomes the live agent ("Workflow published successfully").

You must save before you can publish; if you try to publish with unsaved edits you'll be prompted to "Save the workflow before publishing."

Version history. Click the clock icon to open the version panel. Each save creates a version; you can select a version to preview it (read-only), and Restore an older version to bring it back as your current draft (then save to keep it). When previewing a historical version the header shows a blue "Viewing v… — Read only" banner and a Back to Draft button.

Other ⋯ actions:

• View Runs — jump to this agent's full call history.
• Duplicate Workflow — copy this agent and open the copy.
• Download Workflow — export the agent as a .json file (this is what you later Import).
• Copy Agent UUID — copy the agent's unique identifier to your clipboard.

5. The Agent section (the heart of configuration)

In Simple view, the left sidebar lists the configuration sections. The first and most important is Agent, subtitled "Define your agent's persona, voice, and starting message." It has two columns: prompts on the left, voice/model settings on the right.

5.1 Naming your agent

Rename the agent at any time by clicking its name in the editor header (a pencil appears) and typing a new name, then pressing Enter. A clear name helps you find it later — many teams include Inbound/Outbound in the name, which also drives the little type hint on the list.

5.2 System prompt

The System Prompt is the most important field — it's the agent's persona, goals and rules, written in plain English. This is where you tell the agent who it is, what it's trying to achieve, and how it should behave.

A good system prompt typically uses simple headed sections, for example:

Tips:

• Be concrete about the goal. "Book a viewing" beats "be helpful".
• Keep replies short. Phone conversations work best in 1–2 sentences.
• List firm rules the agent must always follow.
• Inject live data by typing {{variable}} anywhere in the prompt — see Variables (section 9) and Memory (section 7). A reminder under the box says: Type {{variable}} to inject data.

At the bottom of the system prompt box there's a timezone picker (globe icon). Set it to the timezone your agent should assume for times and dates (e.g. "London"). Leave it on Auto to use the system default.

5.3 First message (the greeting)

The First Message is exactly what the agent says the moment the call connects — for example "Hi, thanks for calling Acme — how can I help you today?". Like the system prompt, it supports {{variable}} injection.

Leave it blank to have the agent wait for the caller to speak first. The hint reads: "Leave blank to wait for the caller to speak first." When a primary language is set, a small flag and language label appear in the corner of the box.

5.4 Call Start Settings (advanced opener options)

Below the First Message is a collapsible Call Start Settings panel with three options:

6. Voice, model and language settings

The right-hand column of the Agent section is where you choose how the agent sounds and thinks. It's split into labelled blocks: Voice Engine, Intelligence, Transcriber, Language, Voice Tuning, Memory and Quick Settings.

If no AI providers are connected yet, this panel shows "No AI providers configured" with a link to set up AI Models. Once providers are connected, the dropdowns fill with available options.

6.1 Voice selection (Voice Engine)

The Voice Engine block has two parts:

1. The voice model — choose a text-to-speech provider/model from the dropdown (e.g. an OpenAI, Deepgram, Rime, Cartesia, xAI or ElevenLabs voice engine). The dropdown groups models by provider and is searchable.
2. The specific voice — once a model is selected, pick the actual voice. Depending on the provider you'll either:
   • choose from a list of built-in voices (each with a short hint about its character), or
   • pick from your voice library using a voice picker (for providers that support custom/cloned voices).

Built-in voice examples include OpenAI voices like Alloy (neutral, balanced), Nova (energetic, bright) and Onyx (deep, authoritative); Deepgram voices like Asteria and Luna; and others. The exact list depends on which providers are connected to your account.

Tip: Changing the voice engine resets the selected voice (because each engine has its own voices), so pick the engine first, then the voice.

6.2 Model selection (Intelligence and Transcriber)

• Intelligence — the LLM that powers the agent's thinking and replies. Choose from the model dropdown. This is the biggest single factor in how "smart" and capable the agent feels.
• Transcriber — the STT model that turns the caller's speech into text. A good transcriber improves how accurately the agent understands callers, especially with accents or background noise.

Both use the same dropdown style as the voice engine: a searchable list grouped by provider. Each model may show a small default marker. You can also pick Account default to simply use whatever your account is configured to use — handy if you'd rather not choose specific models.

6.3 Language

The Language block sets the spoken language(s):

• Primary language — choose from a flagged list (English US/UK, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Arabic, Japanese, Chinese, Korean, Hindi, Turkish, Russian, and more). Choose Auto-detect to let the system decide.
• Additional languages — for voice engines that support multiple languages, you can add extra languages so the agent can switch mid-conversation. Selected extras show as removable chips.

If the chosen voice engine only supports English, you'll see "This voice engine supports English only." and the additional-languages option is hidden. The set of available languages adapts automatically to the voice engine you picked.

6.4 Quick Settings

Two handy toggles sit at the bottom of the panel:

These mirror settings also found in the Behavior section (section 8).

7. Voice tuning controls

The Voice Tuning block lets you fine-tune how the voice sounds and how the conversation paces. The available sliders depend on the voice provider you chose — the platform only shows controls that provider supports.

7.1 Voice parameter sliders

If a provider has no tunable parameters, you'll see "No tuning parameters available for this provider."

7.2 Pipeline controls (all providers)

Below the voice sliders are two conversation-pacing controls that work with any provider:

8. Caller memory

If your plan includes it, a Memory block appears in the Agent panel. Caller memory lets your agent recognise repeat callers and remember their details between calls — so a returning customer gets a personal, "you again, welcome back" experience instead of starting from scratch.

8.1 Turning memory on

In Simple view the Memory block shows a compact Caller Memory toggle:

• On — "Callers recognised & remembered"
• Off — "Memory disabled"

The fuller memory configuration (available in the agent's advanced settings) lets you control exactly what is remembered:

You can also set the History depth (how many recent calls to summarise, 1–20).

Privacy: Payment references store tokenised references only — never raw card numbers. Memory is designed to be helpful without holding sensitive payment data.

8.2 Using memory in your prompt

Memory only changes the conversation if you reference it in your system prompt. The settings panel provides copy-ready template variables you can paste in:

A typical use is to paste {{caller_memory}} near the top of your prompt and add a line like "If {{caller_known}} is true, greet {{caller_name}} warmly by name." Telephone calls are supported automatically.

9. The other Simple view sections

Beyond the Agent section, the sidebar has six more sections. Here's what each does.

9.1 Tools

The Tools section — "Manage system capabilities and custom integrations for this agent." — controls the actions your agent can take during a call. There are two groups:

System Tools (built in):

To set up a tool, click its row to open a side panel, fill in any fields, then click Enable Tool. Each tool has a switch to turn it on or off for this agent. Tools that support multiple destinations (like Transfer Call) can hold several configurations — each gets its own sub-toggle.

Custom Tools (your own integrations): These let the agent call your own systems over the web. Click + Add webhook to create one:

The agent calls this URL during conversations, sending parameters as JSON. You can refine headers and schema on the dedicated Tools page after creating it. Existing custom tools are listed with a search box and an on/off switch each.

Note: Tools are shared resources for your account. Enabling a tool here switches it on for this agent; deleting a tool configuration removes it everywhere.

9.2 Knowledge

The Knowledge section — "Attach documents your agent can retrieve during conversations." — lets you give the agent reference material it can pull facts from mid-call (price lists, FAQs, policy documents, etc.).

• Documents you've uploaded appear in an All Documents list, each with a file size and a processing status (completed, pending/processing, or failed). A document can only be attached once it's completed.
• Flip a document's switch on to attach it to this agent. Attached documents show as removable chips at the top.
• Some documents are marked global (shared across agents).
• Use the search box to find a document by name.

Documents themselves are uploaded in the Knowledge Base area of the platform; here you simply choose which ones this agent uses.

9.3 Behavior

The Behavior section — "Call limits, turn detection, voicemail, and conversation settings." — groups the agent's call-handling rules into cards:

9.4 Variables

The Variables section — "Inject dynamic data into prompts and track call outcomes." — has two cards:

• Template Variables — define {{variable}} placeholders and their default values. Use them in your system prompt and first message to inject data. For example a variable business_name with value Acme Lettings lets you write {{business_name}} in the prompt.
• Call Dispositions — define outcome codes (e.g. key SALE, label "Sale made") that get tracked for analytics, so you can report on how calls ended.

Add an entry by typing a key and value and clicking +. Edit values inline, and remove entries with the trash icon.

9.5 Phone

The Phone section — "Assign inbound phone numbers to this agent." — connects real phone numbers so callers can reach the agent. It lists:

• Assigned to this agent — numbers currently routing inbound calls here (remove with the ✕).
• Available to assign — unassigned numbers you can Assign with one click.
• Assigned to other agents — numbers used elsewhere, which you can Reassign to this agent if needed.

If you have no numbers yet, you'll see "No phone numbers configured" with a note to add them in the Phone Numbers area first. A compact version of this assignment control also appears as a banner at the top of the Simple view.

Even without a phone number, you can always talk to your agent through the web call tester or the embeddable widget (sections 10 and 11).

9.6 Conversations

The Conversations section is this agent's call history. It shows a paginated table (20 per page) of past calls with:

Click any row to open a detail panel with the call's recording and transcript (when available). Use the refresh icon to reload, and Previous/Next to page through history.

10. The embeddable web call widget

The Widget section — "Embed this agent as a voice widget on any website." — turns your agent into a floating "Talk to us" button you can drop onto your own website, so visitors can have a voice conversation with your agent right from their browser (no phone call needed).

10.1 Publishing the widget

At the top is a Published toggle ("Allow embedding on external websites"). Turn it on to make the widget live — you'll see a green "Widget is live" indicator, and a unique Widget Key is generated automatically. Until you publish, the embed code and live preview are unavailable.

10.2 Choosing a mode

10.3 Appearance (Hosted UI)

For the hosted widget you can style the button:

A live, interactive Preview shows exactly how the widget will look and behave, with a Desktop/Mobile toggle. Once published, you can click the preview button to test a real call right inside the editor.

10.4 Embed code and identifiers

When published, the Embed Code block gives you a ready-to-paste snippet:

• Hosted UI — a small HTML <script> snippet to drop into your site that loads the floating widget.
• Headless / SDK — a JavaScript snippet using the Sysevo SDK, plus a link to the full developer reference.

The Identifiers block shows your Agent ID and Widget Key with copy buttons. You can Regen (regenerate) the Widget Key if you ever need to rotate it — note this invalidates the old key, so update your embed code afterwards. Every hosted widget shows a small "Powered by Sysevo" label.

<script>
  window.SysevoWidget = {
    agentId:     "a1b2c3d4-...",
    widgetKey:   "9f8e7d6c-...",
    theme:       { color: "#6366f1" },
    position:    "bottom-right",
    buttonLabel: "Talk to us"
  };
</script>
<script src="https://cdn.sysevo.com/widget/v1/embed.js" async></script>

11. Testing your agent

You never have to publish to find out whether an agent works — there are several ways to try it live.

11.1 The Test panel (in the editor)

Click Test in the editor header to open a tester panel on the right. It lets you place a web call to the agent directly from your browser: you'll be asked for microphone permission, then you can speak to the agent and watch the live transcript appear. This is the fastest way to iterate — tweak the prompt, save, and test again.

The tester needs the agent to be saved first. If you place a test call with unsaved edits, save them so the test reflects your latest changes.

11.2 The Call action (from the list)

On the Your Agents page, hover a row and click Call to open the same kind of web-call tester in a slide-out panel — handy for a quick check without opening the full editor.

11.3 Web Call experience

A web call shows a status (Ready → Connecting → Connected/Live), an animated waveform while you're connected, and a running transcript of both sides of the conversation. You can choose your microphone (if you have more than one), Mute/Unmute, End Call, start a New Session, or Retry if a connection fails. The transcript also shows when the agent moves between conversation steps.

11.4 Test Call (real phone)

If a phone number is assigned to the agent, the editor also shows a Test Call button that places a real telephone call so you can experience the agent exactly as your customers will.

12. Common workflows

12.1 Launch your first agent (fastest path)

1. On Your Agents, click New Agent.
2. Choose Agent Builder and work through the wizard (Industry → … → Review).
3. When it opens in the editor, click Test and have a quick conversation.
4. Adjust the System Prompt and First Message in the Agent section if needed.
5. In Phone, assign a phone number (or set up the Widget for your website).
6. Click Save, then Publish. Your agent is live.

12.2 Build a variation of an existing agent

1. On the agent's row, open the ⋯ menu and choose Duplicate (or use Duplicate Workflow in the editor).
2. Rename the copy clearly.
3. Change what differs (prompt, voice, language, tools).
4. Save and Publish.

12.3 Make your agent multilingual

1. In the Agent section, pick a Voice Engine that supports multiple languages.
2. Set the Primary language, then add Additional languages.
3. In your System Prompt, tell the agent it may switch languages to match the caller.
4. Test with a call in each language, then Save and Publish.

12.4 Recognise returning customers

1. Ensure your plan includes memory, then turn on Caller Memory in the Agent section.
2. Configure what to remember in the memory settings (name, history, preferences…).
3. Paste {{caller_memory}} into your system prompt and add a line greeting known callers by name.
4. Save, Publish, and test by calling twice from the same number.

12.5 Put a voice agent on your website

1. Open the agent and go to the Widget section.
2. Turn on Published, choose Hosted UI, and style the button (colour, position, label).
3. Use the live Preview to test a call.
4. Copy the Embed Code and paste it into your website's HTML.

13. Tips and good practice

• Save often, publish deliberately. Saving creates a draft you can iterate on safely; publishing pushes changes to live callers. The yellow Unsaved changes badge is your reminder.
• Keep replies short. Instruct the agent to answer in 1–2 sentences — long monologues feel robotic on the phone.
• Use the dictionary (Behavior → Transcription Dictionary) for brand names and unusual words so the agent hears them correctly.
• Match interruption sensitivity to your audience. Chatty consumer lines often suit High; agents that must read out important details may suit Low.
• Set a sensible Max Call Duration to avoid runaway calls, and an Idle Timeout so the agent doesn't sit on dead air.
• Test in the browser first, then with a real Test Call once a number is assigned.
• Archive instead of delete when retiring an agent you might want back — deletion is permanent.
• Version history is your safety net — if a change makes things worse, restore an earlier version.

14. Troubleshooting

Looking for related guides? See the documentation for Phone Numbers, Knowledge Base, Tools, Campaigns, and Voice Analytics for deeper coverage of the areas this guide references.

Choosing an AI model

Each agent runs on a large language model you can choose per agent. Sysevo is model-agnostic and supports leading providers, so you can balance quality, speed, and cost:

The model is bundled into your all-in minute rate — there's no separate per-token bill to manage. Pick a faster, lighter model for simple flows and a stronger model where the agent needs to reason.

Speech: how the agent hears and speaks

• Speech-to-text (hearing): Deepgram Nova-2 / Nova-3, AssemblyAI Universal-2, Whisper Large v3, or Azure.
• Text-to-speech (speaking): ElevenLabs Turbo v2.5, PlayHT 3.0 Turbo, Cartesia Sonic, or Azure Neural voices.

These combine to deliver under 800ms to first word, so conversations feel natural rather than walkie-talkie. Agents support 30+ languages — one agent can greet and converse in a caller's language without you hiring multilingual staff.

Custom (cloned) voices

Beyond the stock voices, you can train a custom voice that sounds like your brand or a specific team member in the Voice Library, then assign it to any agent for consistent brand identity across every call.

Caller Memory in depth (Growth and above)

Caller Memory is what makes a Sysevo agent feel like it already knows your customers. Every returning caller is recognised by phone number, and their profile is loaded before the agent speaks — zero re-introduction, full context.

How it works — three phases

1. Pre-call recall (under ~100ms). When a call arrives, the number is matched against your contact database and the caller's name, preferences, and history are injected into the agent's context before it greets them. (Demonstrated figures: ~43ms average recall, 98.6% match rate.)
2. Real-time extraction. During the conversation the system listens and extracts discrete facts — preferences, contact details, decisions — tagging each with a confidence score and a data-sensitivity flag.
3. Post-call consolidation. After the call, profiles update automatically: new facts are added, corrections supersede old data, and a plain-language summary is stored for next time.

What gets remembered

Full and preferred names · email and physical addresses · appointment and service preferences · complete call history with summaries · logged complaints and outstanding issues · payment references (sensitive, opt-in) · and any custom domain-specific facts you define.

Memory survives agent updates — re-publishing or editing an agent never wipes what it knows about your callers. Profiles compound over time: names and history in month 1, anticipatory preferences by month 3, deep contextual profiles by month 12.

Privacy & compliance

• GDPR / UK GDPR compliant by design, with right to erasure (verbal or API-triggered) and audit logging.
• Configurable retention — 30, 90, 365 days, or custom.
• Sensitive-data flagging for payments, allergies, complaints, etc.
• Data residency options (EU, UK, US).
• Do-Not-Call flags propagate to outbound campaigns automatically.