How multi-agents agents work
What a multi-agent is
A multi-agent brings multiple specialized child agents into one chat interface - a single entry point with a dropdown to pick the right agent.
The multi-agent has no AI configuration. All AI behavior- persona, intelligence, data sources, tools, and actions- lives in each child agent. The multi-agent holds the interface together.
How responsibilities are split
| Controlled by multi-agent | Controlled by child agent |
|---|---|
| Theme, colors, logo, and layout | Persona and conversation style |
| Header text and welcome message | Intelligence settings |
| Conversation UI behavior | Data sources and documents |
| Public/private status | Actions and tools |
| User feedback, sharing, and export settings | API and MCP access |
| Citation Settings, Citation Types, and PDF Citations | Citations Advanced Settings (Knowledge Base Visibility, Full Document Reading, Numeric Search Optimization) | |
Multi-agent settings apply across the entire session regardless of which child agent is selected, keeping the interface consistent while each child agent stays tuned for its purpose.
Citation Settings, Citation Types, and PDF Citations always use the Multi-Agent's configuration, overriding whatever each child agent has set - this keeps citations consistent no matter which child agent answers. Citations Advanced Settings - Knowledge Base Visibility, Full Document Reading, and Numeric Search Optimization - are the exception: those still come from each child agent. PDF Viewer works the same way on a Multi-Agent as on a regular agent - clicking a PDF citation opens the file in the in-chat viewer - and older Multi-Agent messages still show their citations correctly, as long as the citation data was already attached to that message when it was sent.
What end-users see
When a user opens a multi-agent with at least one child agent attached:
- Interface follows the multi-agent's color scheme and branding
- Dropdown in the top left lets the user pick a child agent
- Each child agent's avatar appears in the dropdown and next to its message bubbles
- Last selection is remembered in the browser and persists across sessions on the same device
If no child agent is pre-selected, the user sees the selector with no prompt input and must choose an agent before chatting.
Users can switch agents at any point during a session. The conversation continues with the same context - nothing is lost when a different agent takes over, whether the user switched manually or Smart Routing handed off automatically.
When a child agent is private
Child agents set to private don't appear in the dropdown when the multi-agent is deployed publicly. Only child agents set to public are available to end-users.
How the multi-agent sits in your account
- The multi-agent appears in the sidebar with its own icon
- On the Agents page, the multi-agent card shows a "Multi-agent" tag, its connected agent count.
- Query data is available per child agent in Build > Agents. The multi-agent does not have its own query count
- The same child agent can belong to multiple multi-agents
- Child agents remain independently deployable- attaching to a multi-agent doesn't change how a child agent works on its own.
- Any changes made to a child agent are reflected universally- whether it is accessed through the multi-agent or deployed independently.
What happens if your plan changes
If Multi-Agents is disabled (e.g. after a plan downgrade):
- Multi-agent cards on the Agents page are greyed out with a lock icon
- Clicking a locked card shows a popup to contact our sales or delete the agent
- Navigating directly to a locked multi-agent's URL redirects to the Dashboard
- Deployed multi-agent behave as if they have no child agents- no prompt input, no dropdown, users cannot send messages
Existing agents cannot be converted into a Multi-Agent. To use the Multi-Agent feature, create a new agent and select Multi-Agent at setup, then attach your existing agents to it.
Related articles
Updated 11 days ago