Multi-agent conversational chat

Multi-agent conversational chat turns your AI agent flow into an assistant that talks to real people — from a widget on your website, from Telegram or from WhatsApp — keeping per-user memory, handing off the conversation to the right specialist and, when needed, verifying identity before sharing sensitive data.

It is the conversational evolution of modular AI agents: the same pieces (brain, tools, RAG, skills), now orchestrated as a team that staffs a chat.

In one sentence

Connect several agents to the same flow, mark one as the default, and each one handles its topic. If the user asks about a different domain, the agent hands off to the right teammate — and everything stays in the same conversation, with the same memory.

The five capabilities

Team with hand-off

Each AI Agent node in the flow is a specialist (Sales, Workshop, Admin…). One is the default and takes the first message; if the request belongs to another domain, it hands off to the right teammate.

Memory per user and channel

Each person has their own session keyed by (client, agent, channel, external id). A rolling summary of the conversation plus a long-term memory with the fields you choose are stored.

Identity verification (OTP)

You can require the user to verify their email with a one-time code before a tool returns sensitive data (invoices, order status…).

Tools over real data

Beyond classic tools, an agent can query a DataStore and answer with real records — automatically filtering to the verified user’s own data.

Web publishing + embed

From the chat trigger you get the chat’s public URL and a snippet to embed it (floating widget or iframe) in WordPress, a store or any HTML page.

Building the team

A team is several AI Agent nodes in the same flow, each with its brain, tools and skills. The only thing new compared to a standalone agent:

1. Give each agent a name and a purpose. On each AI Agent, the Agent name and Purpose fields describe what it handles (for example “Sales — pricing, availability and quotes”). That purpose is what the team uses to decide who to hand off to.

2. Mark the default agent. Enable Default agent on one of them (shown with a ★ in the editor). It answers the first message of every conversation.

3. Connect the chat trigger. The Chat Trigger node passes the user message to the team and returns the active agent’s reply.

How is the hand-off decided?

No rules to wire. On every turn, the active agent receives the list of teammates with their purpose and a handoff_to_<agent> tool. If it detects the request isn’t its own, it calls it and the turn continues with the correct specialist. To avoid infinite ping-pong there is a maximum of 2 hand-offs per turn.

The active agent is remembered in the session: if Sales handed off to Workshop, that user’s next messages keep being handled by Workshop until the topic changes.

Per-user memory

Memory works on two levels, with no admin required:

• Rolling summary — as the chat grows, it is condensed into a summary that travels on every turn. The agent “remembers” what was said without dragging the whole transcript.
• Long-term memory — specific user facts (name, company, preferences, verified identity…) that persist across conversations.

Choose what to remember

On the AI Agent you can configure the memory fields (for example name, company, taxId). The agent fills them in during the chat and they become available in the following turns and sessions for that same person.

Identity verification (OTP)

When a tool returns sensitive information (invoices, an order’s status, account data), you can require the user to prove they are the account holder before answering.

1. Mark the tool as protected. On the Tool node enable Requires verification.

2. The agent asks for a contact. While the user is not verified, the protected tool stays hidden and the agent offers to verify identity.

3. A code is emailed. Floogos sends an OTP to the given address using the same email template as the security OTPs for forms and passwords.

4. The user enters the code. On confirmation, the identity is verified for that session and the agent answers in the same turn — no waiting.

Data isolation

Once verified, data tools automatically filter to that email’s records. A verified user never sees another person’s invoices or orders.

Tools that query real data

So the agent answers with data and does not make things up, connect tools that query real sources. The most direct way is a DataStore query tool: you point it to the data group and the tool searches there (for example, “find invoice by number”).

• If the tool is protected, the search is scoped to the verified email’s records.
• You can connect several tools to the same agent (find invoice, order status, create ticket…). The model picks which one to use per request.

Wrapping these tools in a Skill (for example “Billing”) makes the agent activate them only when the request is relevant.

Publishing the chat

From the Chat Trigger you get everything you need to go live:

Public URL

A /(...) link ready to share or open in the browser. You can rename or regenerate the URL token whenever you want.

Floating widget (embed)

A <script> snippet that adds a chat button in the corner of your site (WordPress, store, landing…). Served from /agents/embed.js.

Iframe

An <iframe> to embed the chat as a panel inside an existing page.

Same team, multiple channels

Web chat and messaging channels share the same team and the same memory. To take it to Telegram, WhatsApp or Instagram, the flow uses the agent router node; each person keeps their session per channel. See Agent Router.

Quick recipe: support with verified billing

1. Chat Trigger → connect the team.
2. “Sales” agent (default): answers pricing and availability; RAG with your catalog.
3. “Admin” agent: purpose “invoices, payments and account status”.
4. On Admin, a “Billing” Skill with a find_invoice tool (DataStore query) marked Requires verification.
5. Configure the memory fields (name, email, taxId).
6. Publish the chat with the embed on your site.

Result: the user enters via Sales; if they ask about an invoice, the team hands off to Admin, which asks to verify the email, sends the OTP, and on confirmation returns that holder’s real invoice — remembering their details on the next visit.

Keep reading

• Modular AI agents overview
• Tool · Skill · Agent Router
• Common problems
• Agent library — ready-made skills, tools, prompts and kits to insert into your agent.