Configuring an Agent

Set up a Limio agent in the admin console: instructions, presentation, tools, knowledge, and publishing.

Agents are configured in the Limio admin under AI → Agents. Each agent has instructions, a presentation, a set of tools, optional knowledge sources, and outcomes. You edit a draft, test it in place, and publish a version when it is ready. Nothing changes for visitors until you publish.

Creating an agent

An agent's name is a short identifier: lowercase letters, digits, hyphens, and underscores. It cannot be changed after creation. The agent named default is the one the chat widget uses.

Instructions

The Details tab holds the agent's instructions: who it is, what it sells or supports, and how it should behave. Write them the way you would brief a new team member. A typical structure:

• Who the agent is and what the shop sells.
• How to qualify a visitor before recommending anything.
• How to handle common objections.
• When to hand off to a human.

You do not need to tell the agent how to behave safely with tools. When an agent has tools, the platform adds its own ground rules around your instructions: stay in character, never expose internal mechanics, only quote offers and prices a tool returned, and present checkout links as clickable links.

Presentation

The Presentation tab controls how the chat surface looks before and during a conversation.

Setting	What it does
Chat title	Heading shown above the conversation.
Greeting	Opening message the agent sends when a conversation starts. Leave blank for the agent to stay silent until the visitor speaks.
Composer placeholder	Hint text in the message box.
Pointers	Up to four starter questions shown as clickable chips beneath the greeting. They disappear once the conversation starts.
Footer	Optional markdown shown beneath the message box.
Theme	Brand colours and bubble style for this agent's chat, with a live preview. See Theming an Agent.

Tools

Tools let the agent act rather than just talk. Attach them on the Tools tab.

Tool	What the agent can do with it
Fetch offers	Read the offers carrying a given offer label, so recommendations quote real names and prices. Configure the label to scope which offers the agent sees.
Build checkout link	Build a checkout link for the offer the shopper chose, prefilled with details collected in conversation (email, name, company) and tagged utm_source=limioagent so agent-driven purchases are attributable. If the agent's reply ever omits a link it built, the platform appends it, so the shopper always receives it.
Get order items	Read what the shopper bought, for use on the order-complete page.
Documents	Request documents from the visitor. An upload control appears in the chat; configure which documents the agent may ask for.

Lead capture is deliberately not in this list: the agent cannot write a lead on its own initiative. Capture runs through outcomes, where you control exactly when it fires and which fields it records.

Knowledge

The Knowledge tab attaches document sources to the agent. When a visitor asks something a document covers, the agent reads the document and answers from it instead of guessing. Attached sources appear as a catalogue the agent can consult; it never reveals to the visitor that it is reading documents.

Outcomes

Outcomes are the goals the agent works towards in every conversation: qualify and send a checkout link, capture a lead, hand off to sales. They have their own tab and their own page: see Outcomes.

Advanced

Two per-reply caps, both optional:

Setting	Default	What it caps
Max output tokens	4096	The length of a single reply.
Max agent steps	8	Tool-use steps within a single reply.

Most agents do not need to change either.

Stream replies

Stream replies is off by default. Turn it on and the reply appears as the agent writes it, a few words at a time, rather than arriving whole once the agent has finished. Streaming changes when the first words appear, not when the reply completes.

It does not make the agent faster. An agent that calls tools spends most of its turn on them and cannot write the answer until they return, so the pause before the text starts is as long as it ever was.

The agent's own words before a tool call are never shown. A model tends to narrate ("let me check our offers") before it calls a tool, and that text is discarded rather than typed out and then snatched back. Only the answer streams.

Streaming also needs a model that supports it. If you turn the toggle on and replies still arrive whole, the model behind your tenant cannot stream: ask Limio, rather than looking for another setting on the agent.

The toggle is part of the agent, so Save and Publish latest apply as they do to any other change. Once published, new conversations pick it up within a minute. A visitor with the page already open keeps the behaviour they loaded with until they reload.

Editing as YAML

The YAML tab shows the whole agent as one document: instructions, presentation, tools, and outcomes together. Problems are flagged as you type, Apply to form loads your edits back into the tabs, and Save persists whatever the form holds. Use it to review an agent in one place or to copy a configuration between environments.

description: Inbound sales agent
systemPrompt: >-
  You are an inbound sales agent on the pricing page...
presentation:
  title: Sales
  greeting: Hi! What brings you here today?
toolsets:
  - name: fetch_offers
    config:
      label: pricing-page
outcomes:
  - id: qualified_send_link
    label: Qualified - Send checkout link
    criteria: The shopper is ready to buy one of the offers you can sell.
    whatToDo: Use @build_checkout_link and present the link in your reply.
    tools:
      - id: build_checkout_link
        tool: build_checkout_link

Testing with Try it

Try it opens a test conversation against your current draft, unsaved edits included. The full agent runs: tools fetch real offers and build real links. Outcome follow-up actions are the one exception: they are reported in the transcript instead of executed, so testing never creates a lead.

Versions and publishing

Publishing an agent makes that version the live one, but it does not put the agent on a site. A workflow is what routes a domain to an agent. Until one does, chat on that domain will not start.

Save creates a new draft version; the live agent is untouched. When the draft is ready, Publish latest puts it live. The Versions tab lists the history, and any version can be published, so rolling back is publishing the version that worked. Unpublishing the live version takes the agent offline.