Agents

Agents are LLM-driven workspace resources. You define a task prompt, attach tools and knowledge, set guardrails and channel behavior, then run or preview the agent in real time.

When to use

• Build a conversational assistant for chat, voice, or general (tool-callable) use
• Ground answers with knowledge bases, tools, and structured arguments
• Keep behavior constrained with guardrails, context limits, and human approval on tools

Create an agent

Open Build → Agents and create a new resource.

1. Name (min 2 characters), optional Description.
2. Channel — how the agent runs:

After create, Vettero opens the agent builder. You can also open an existing agent from the Agents list.

Open the builder

The builder has a left activity bar, a main panel, and (on Config) a Preview pane on the right. Use Save in the header when the config is valid.

Activity bar

Write the task and input

On Config, above the accordion:

Write clear identity, allowed topics, which tools to call, and when to hand over or escalate. For prompt, LLM, and settings tips, see Prompts.

LiquidJS and {{ … }}

Task and Input support full Liquid templates (output, tags, filters) against the run context. Autocomplete suggests schema fields when you type {{. See Variables for Liquid syntax, filters, tags, env vars, and path bindings.

# tool mentions

Type # in the Task editor to pick a tool from the same tree as Manage Tools (blocks, agents, workflows, functions, skills, APIs, MCP, knowledge bases, queries, service tools, chat widgets, and so on).

• Inserts a # mention in the prompt so the model is told about that tool.
• Selecting a tool mention adds it to the Tools list if it is not already attached.
• Use # when the agent should be able to call that capability.

@ resource mentions

You do not know workspace resource ids. Type @ for autocomplete and pick a resource by name (not the same as attaching a callable tool).

• You keep seeing the name in the editor. The stored token is the id (@type:id, for example @kb:…, @chat_profile:…), which is what the agent needs (for example to pass into a tool call).
• Use @ so the prompt can point at a concrete resource (handover profile, KB, table, person) without you copying ids.
• Table, chat profile, user, and role mentions are prompt references only and are not auto-added as tools. Attach callable tools with # or the Tools tab.

Channel visibility

Which accordion sections appear depends on channel and whether the agent is conversational. On the main agent builder, a channel agent is treated as conversational.

General is a functional (tool-callable) agent — Suggestions apply to Chat only.

Configure Basic

Identity for the assistant:

On the main Agent page, name and description may be read-only in the builder — edit them on the Agent details page instead.

Configure the LLM

Powers replies (or realtime audio).

Text / chat / voice (non-realtime)

Realtime Voice

Uses a separate realtime config instead of the text LLM:

Changing provider clears model/voice; changing model clears voice.

Configure Voice Config

Shown for Voice and Realtime Voice only.

STT / TTS (Voice channel)

Required before voice preview works:

Realtime Voice hides STT/TTS (the realtime model handles audio).

Call behavior (both voice types)

Configure RAG

Conversational agents only. Attach a knowledge base so the agent can retrieve grounded content.

If the selected model lacks tool-calling, strategies remain editable but retrieval may be unreliable for agentic modes. You can also mention a knowledge base in the Task with @kb:… or attach it as a tool on the Tools tab.

Configure Tools

Select tools the agent can call. The model must support tool-calling; otherwise the tools UI is blocked with a warning.

Use Manage Tools to pick callable tools, grouped for example as:

Typing # in Task and selecting a tool also attaches it to this list.

Tool approval

When a tool requires approval, the agent pauses the chat/run until someone approves or rejects. On reject (or non-approval), the tool is not executed and the model is told the outcome.

For workflow Request Approval / Request Information blocks and the Approvals UI, see Approval Flow.

Configure Guardrails

Guardrails are on by default. The Policy is added to the task. Put {{ ctx.rejectionPolicy }} in the task to choose where it goes. If you leave it out, the policy is appended.

In a live chat the user sees the reply until the checker rejects it. The text is then replaced with the fallback.

The policy is still added on voice and realtime voice. The checker does not run there.

Configure Arguments

Optional JSON Schema for structured inputs into the run.

Use Liquid {{ args.fieldName }} (and other rendered variables) in Task and Input. In Preview, Edit Args fills values for the session.

Configure Result

Hidden for Realtime Voice. When enabled, the agent returns a structured JSON result instead of plain text.

Configure Suggestions

Chat channel only (not General — General is a functional agent). Also hidden for voice, realtime voice, and WhatsApp. Suggests follow-up replies during conversations.

Configure Starters

Up to 5 prompts shown as buttons when a chat starts with no messages. The user taps one to send it as their first message. Hidden for voice, WhatsApp, and non-conversational setups. Add or remove starter message strings in the list.

Configure Context

Conversational text channels only (hidden for voice / realtime voice). Controls how long histories are kept.

Configure STM

Conversational text channels only (hidden for voice / realtime voice). STM is Short-Term Memory for files (and vision when applicable).

Preview and debug

The right pane tests the live configuration.

Fix validation errors from the Config panel before preview will start.

Where agents are used

Limits and tips

• Pick a model with tool-calling if you need tools, agentic RAG, or Load Files into STM.
• Structured result and suggestion LLMs need structured-output support.
• The guardrail policy is part of the task. The optional checker is a separate LLM or TypeSafe call and adds cost.
• Starters max out at 5 agent-opening messages.
• Tool approvals pause the run until approve/reject; role holders use Approvals (or in-chat when they hold an approval role). See Approval Flow.
• Voice preview needs STT/TTS (Voice) or provider + model + voice (Realtime Voice) before a voice session can start.
• Channel choice at create time drives which Config sections you see; change channel-related behavior carefully when switching surfaces.