A AtomicX Chat
Help Sign in

Updated October 5, 2026

AI agents

Overview

An AI agent defines how the AI behaves in your chats: its role, its tone and the rules it follows. For example, you might have a friendly "Sales Assistant" and a strict, fact-only "Support Agent". Visitors don't choose an agent. You attach one to an AI Chat node in a workflow, and it's applied automatically.

Agents have versions. When you Publish an agent, a new numbered version is saved and becomes the one your chats use. Only published agents can be selected in a workflow.

You'll find agents under AI → AI agents.

Steps

Create an agent

  1. Go to AI → AI agents and click Create agent.
  2. If templates are available, you can pick one from Start from Template to fill in the form for you. You can still change every field.
  3. Enter a Name, for example "Customer Support Agent". This field is required.
  4. Optionally, add a Description of the agent's purpose.
  5. Write the Agent Instructions (System Prompt). Describe the agent's role, tone and rules. For example: "Only answer using the provided knowledge base. Be professional and friendly. Ask a clarifying question when unsure. Never make up information." Click Improve with AI to have AI rewrite your instructions with a clearer role, tone and rules while keeping your specific facts. If the box is empty, it drafts instructions from the agent's name and description. The result replaces the text in the box, so review it before you save.
  6. Choose an AI model: Regular model or Premium model. Under it, Which model should I use? describes each one; click a model there to select it.
    • The Regular model is fast and cost-efficient, and the right choice for most sales and support chats.
    • The Premium model thinks before it answers, so it handles complex, multi-step questions better. It uses more credits and replies more slowly.
    • Each model shows roughly how many AI credits it uses per message, for example "≈ 0.6 credits per message". See AI credits.
    • Choosing a model fills in Max Output Tokens for you: 1,000 for the regular model and 4,000 for the premium model, which uses part of it for thinking. You can still change it.
  7. Set the Temperature, from 0 to 1. Lower values give safer, more factual answers. Higher values are more conversational. The Quick Guide on the page suggests ranges for different industries.
  8. Check Max Output Tokens, from 100 to 4,096. It's filled in when you choose a model and limits the length of each reply. Higher values allow longer answers but use more AI credits. For the premium model, keep it at 2,000 or more.
  9. Optionally, choose Skills. Skills are extra abilities the AI can use during replies. AtomicX manages the skills catalog, and only the skills that are switched on for your account are listed.
  10. Optionally, choose MCP tools. These let the agent look up or update data in your own systems through connected MCP servers. See Give AI agents MCP tools.
  11. Click Create Agent.

New agents start as Draft.

Publish an agent

  1. On the AI Agents page, find the agent's card.
  2. Click Publish.

The agent's status changes to Active, and a new version is saved. Publish again every time you edit an agent. Your chats keep using the last published version until you do.

Attach an agent to a workflow

  1. Go to Chat → Workflows and open your workflow.
  2. Select an AI Chat node, or add one.
  3. Under AI Configuration, choose your agent from AI Agent. Only published agents appear here.
  4. Optionally, choose a Knowledge Base so the agent can answer from your content.
  5. Click Save.

Offer a workflow during AI chat

Under Edit → Callable workflows, tenant owners and tenant admins can attach a workflow that the agent offers when it fits the conversation. The visitor confirms before it starts, and the agent continues afterward with the collected information. Save the attachment and publish the agent to activate it. See Let an AI agent offer a workflow.

For evidence-based changes, use AI → Improvements to review, evaluate, approve and publish a candidate. See Improve your chat.

Edit, roll back and archive

  • Edit opens the agent so you can change its settings. Click Update Agent to save, then Publish to make the changes live.
  • Versions shows the Version History, with the publish date and settings of each version. Click Set as Current on an older version to go back to it.
  • Archive hides an agent from use after you confirm. Archived agents can't be published. Click Restore to bring an archived agent back.

Import an agent from a web address

  1. Click Import from URL.
  2. Paste a Source URL pointing to a Markdown or JSON document that describes the agent's persona.
  3. Click Import. The AI reads the document, writes the instructions and creates a new draft agent.
  4. Review the instructions on the edit page, then publish.

Shopify customer and staff agents

Shopify settings can create two separate agent drafts for live store retrieval. Review and publish both. Use the customer agent in your website chat; customers must verify order access. The staff agent runs only in the private staff chat for explicitly enabled team members and cannot be assigned to a public widget. Its transient history is not shown in Chat → Conversations. See Shopify integration.

Reply in the visitor's language

In the chat widget, AI replies follow the language the visitor writes in, starting from the website's language. Keep your agent's instructions focused on its role, tone, facts and rules, and test answers in other languages against the same knowledge sources. Earlier messages and human replies retain their original wording.

Your authored workflow messages and callable offer titles are translated automatically into the visitor's language. To review or correct them, use the Title translations and Workflow step translations panels under Callable workflows and choose a language in Translate to:. Republish the agent after changing a callable workflow. See Chat languages and translations.

Tips

  • If an agent's model isn't available, the agent keeps working on your plan's default model. The agent editor and the Agents list show a notice such as "Running on the regular model — the model this agent was set to isn't available." Choose a model and save to clear it.
  • Write specific rules, not vague guidance. "Only answer questions about our products" works better than "be helpful".
  • For regulated industries such as legal, medical or finance, keep Temperature at 0.4 or lower.
  • Test your chat after every publish to check the agent behaves the way you expect.
  • Prebuilt agents are also available. See Agent gallery.
  • Shops, dealers and galleries: add the List products skill so the agent can list what's on your website, with photo cards and prices. See List products in chat.
  • Clubs: add the MemberX: skills so the agent can list events, take RSVPs and look up club information. They're listed only when your plan includes MemberX. See MemberX overview and setup.
  • To check which agent, skills and tools produced a reply, select the conversation in Chat → Conversations and click Show AI Details, or look at Tool executions and Prompts in effect for the whole conversation. See View conversations.
An unhandled error has occurred. Reload ×