VoysseAgents
An agent is the AI persona a client's customers talk to — its instructions, tone, model, generation settings and multimodal capabilities.
An agent is the AI persona a client's customers talk to — its instructions, tone, model, generation settings and multimodal capabilities. Each agent belongs to exactly one client. For the platform overview see the README.
Contents
- Create an agent
- Configure after creation
- Model & parameters reference
- Multimodal capabilities
- Test it live
Create an agent
From Agents → New agent a 5-step wizard walks through the setup. You can go back and forth between steps; nothing is created until the last one.
- Template. Pick an industry starter — Restaurant orders, Real estate leads, Clinic appointments, Online store support or Customer support — or Start from scratch. A template pre-fills the description, main instructions and personality/tone with a ready-made prompt for that use case; you can edit everything afterward.
- Identity. Choose the owning Client and set the agent's Name (required) and Description.
- Instructions. Write the Main instructions (goals, tasks, limits, rules) and the Personality or tone. A progress bar and a live token counter (
~N tokens, estimated at ~4 characters per token) show how much of the model's context window the prompt text uses so far. - Model & parameters. Set the Timezone (injected into the system prompt so the agent knows the local date/time), the Provider (OpenAI or Anthropic — the agency's API key for that provider is used, configured in Settings), the Model, and three generation sliders: Temperature, Max response tokens and Conversation memory (exact ranges and defaults in the reference table below). The context-usage bar reappears here against the selected model's context window.
- Review & create. A summary of name, client, template, provider and model before submitting.
Configure after creation
An agent's detail page (/agents/{id}) has five tabs:
- Details — everything from the wizard, plus a Business brief: optional structured fields (what the business does, products & services, target audience, key info & policies, the agent's main goal, always-do / never-do rules) that compose into the system prompt alongside the free-form instructions. All Details-tab fields, including provider/model/timezone and the three generation sliders, are edited here and saved together with Save configuration. A Publish / Unpublish toggle in the header controls whether the agent is active (an inactive agent doesn't reply and doesn't spend tokens).
- Knowledge — manual context, PDF uploads and Q&A pairs. See Knowledge base.
- Tools — per-agent HTTP tools and MCP servers. See Custom tools.
- Widget — enable/configure the embeddable web chat widget (greeting, color, position) and copy its embed snippet.
- Playground — a live chat against this agent, locked to it.
Model & parameters reference
| Field | Range | Default | Effect |
|---|---|---|---|
| Provider | openai | anthropic | openai | Which AI provider's API is called; uses the agency's key for that provider (configured in Settings). See AI providers. |
| Model | any model string for the chosen provider (free text allowed) | — (blank) | The exact model requested from the provider's chat endpoint (Responses API for OpenAI, Messages API for Anthropic). |
| Timezone | any IANA timezone | the browser's timezone at creation, UTC in the API | Sets the "current date and time" line injected into the system prompt. |
| Temperature | 0.0 – 2.0, step 0.1 | 0.7 | Passed as temperature to the provider. Lower is more predictable/focused, higher is more creative. If the model rejects sampling parameters (e.g. some reasoning models, or Anthropic's temperature <= 1 constraint), the backend retries once without it. |
| Max response tokens | 256 – 8192 in the UI slider (256 step); the API accepts 1 – 32000 | 2048 | Upper bound on the length of each reply (max_output_tokens for OpenAI, max_tokens for Anthropic). |
| Conversation memory | 0 – 100 in the UI slider (step 1); the API accepts 0 – 200 | 30 | How many of the most recent messages in the conversation are sent back to the model as history on each reply. |
The API's Pydantic schema (apps/api/app/schemas.py) enforces slightly wider bounds (max_tokens up to 32000, memory_limit up to 200) than the dashboard sliders expose — useful to know if you're calling the API directly.
Multimodal capabilities
Under Details → Capabilities two independent switches turn on multimodal understanding for incoming messages (WhatsApp, the web widget, and manual uploads in the playground):
- Image recognition — when enabled, an incoming image is sent to a vision model (via the OpenAI Responses API) with an instruction to describe it; the resulting description is prepended to the message text before the agent replies. The model used is the agent's image model setting (default
gpt-4.1, any vision-capable model can be typed in), falling back to the agent's main model if left blank. - Audio recognition — when enabled, an incoming voice note is sent to an OpenAI-compatible
/audio/transcriptionsendpoint; the transcript is prepended to the message text. The model used is the agent's audio model setting (defaultwhisper-1).
Both capabilities always use the agency's OpenAI key, regardless of which provider the agent's chat model is configured with — enabling either capability with no OpenAI key configured means incoming media falls through untranscribed/undescribed at reply time. If a capability is off, that kind of media is ignored (a placeholder is used instead of failing the conversation).
Test it live
Before publishing, or any time after, use the Playground to chat with the agent exactly as an end user would, including tool calls and their results. See Playground walkthrough for the full walkthrough.