Agents

AIML.chat is more than a Q&A chatbot. On paid plans you can turn your assistant into a multi-agent team that routes questions to specialized agents, hands off between them, calls MCP tools to take real actions, and asks for your visitor's confirmation before doing anything that writes.

This page covers what the agents are, how to enable them, and how to configure them.

RAG mode vs Agent mode

RAG mode (default)Agent mode Answers grounded in your contentYesYes Cited sourcesYesYes Specialized agents + handoff—Yes MCP action tools—Yes Safe actions (write-confirm)—Yes PlanAll plansPro and up

In RAG mode every question takes the same grounded-answer path. In Agent mode the orchestrator picks the right agent for each message and lets agents hand off to one another.

The default agent team

The first time agent mode runs, your website is seeded with a default team. You can edit every one of them.

Support — the first responder. Answers from your indexed content and hands off when a question is better served elsewhere. Always available on any agent-mode plan.

Sales — pricing, plans, demos, and lead qualification. *(Agency plan and up.)*

Technical — API, integration, and code questions; can file issues through a connected MCP tool, but only after explicit confirmation. *(Agency plan and up.)*

Custom — one optional agent with a prompt and tools you define. *(Agency plan and up.)*

Each website has exactly one agent per role. The orchestrator routes by intent and confidence, so a visitor never has to pick an agent — they just ask.

Enabling agent mode

Open your website in the dashboard and go to the Agents tab.

Toggle Agent mode on.

That's it — the change is served through your widget config, so you do not need to re-embed the script or change anything on your site. Existing embeds pick up the new mode automatically.

Agent mode is a Pro-and-up feature. On Free/OSS plans the toggle is replaced by an upgrade prompt, and any agent-mode request transparently falls back to RAG.

Editing an agent

In the Agents tab, expand any agent to edit:

Name — shown in the widget when that agent is answering.

System prompt — how the agent behaves. Visitors never see it.

Escalation threshold — a confidence floor from 0 to 1. When the agent's confidence drops below it, the orchestrator hands off to another agent or escalates to a human.

Enabled — turn an agent on or off. Support can be toggled on any agent-mode plan; Sales, Technical, and Custom require the Agency plan or higher.

The built-in Support, Sales, and Technical agents can be disabled but not deleted. The Custom agent can be created and deleted freely.

Handoff

Agents hand off to one another automatically. For example, the Support agent passes a pricing question to Sales, or an API question to Technical. In the widget the visitor sees a small divider when a handoff happens, so the conversation stays transparent. Handoffs and the agent that produced each answer are recorded on the conversation transcript.

Tools (MCP)

Agents can call MCP servers to fetch live data or take actions — look up an order, create a Linear issue, check a build. You install a server once for your account, then link it to the specific agents that should use it.

See the MCP marketplace doc for installing and linking servers. The link controls live on each agent under Tools in the Agents tab.

Safe actions (write-confirm)

Any MCP tool that changes something (creates a ticket, sends a message, issues a refund) is wrapped in a two-turn confirmation:

The agent proposes the action and the widget shows a confirmation card describing exactly what will happen, and which server it targets.

Nothing runs until the visitor approves. On approval the action executes and the result is streamed back; on decline it's dropped.

Read-only tools (search, lookups) run without a prompt. This keeps an agent from taking an irreversible action on a visitor's behalf without a human in the loop.

Testing your agents

The Agents tab includes a live test panel that runs the real widget in agent mode against your site — so you can try handoff, tools, and write-confirm before you enable it for visitors. Test messages use the live pipeline and count toward your monthly quota.

Next steps

MCP marketplace — install action tools and link them to agents

Playground — try a grounded assistant on any site, no signup

Widget customisation — appearance, events, JS API