Agent Mode

Agent Mode transforms a Custom Model into a conversational AI assistant. Instead of a standard prompt-response exchange, the AI Agent interprets natural language requests and executes them — generating charts, querying your semantic layer, managing platform resources, and remembering your preferences across sessions.

Enabling Agent Mode

Open the Custom Model editor, go to the General tab, and switch on Enable Agent Mode.

Once enabled, every session that uses this model runs through the AI Agent. Standard sessions on the same model without Agent Mode are unaffected.

Recommended Provider Model

Agent Mode works with any Chat-type Provider Model. For best results, use a capable model (Claude 3.5 Sonnet, GPT-4o, or equivalent). Smaller or open-weight models often lack the reasoning depth required for reliable chart generation.

What Users Experience

When Agent Mode is active, the conversation interface changes in several ways:

Phase Indicators

For chart generation requests, the Agent shows real-time progress across three phases:

Phase	What is happening
Analyzing	The Agent evaluates your request against the model's data and semantic layer
Generating	Data is queried and the chart dataset is prepared
Visualizing	The chart is rendered and embedded in the conversation

Chart Generation

Users describe what they want to see in plain language:

• "Show me monthly revenue by product category as a bar chart."
• "What are the top 10 customers by total order value?"
• "Compare gross margin across store regions."

The Agent interprets the intent, selects the right dimensions and metrics, runs the query, and returns the chart inline. If clarification is needed — for example, if the request is ambiguous — the Agent asks a follow-up question before proceeding.

After a chart is generated, users can iterate conversationally:

• "Make it a pie chart instead."
• "Filter to only the EMEA region."
• "Break it down by quarter."

Follow-Up Suggestions

After each chart, the Agent generates 2–3 follow-up prompts to help users explore related insights. These appear directly below the chart and can be clicked to continue the analysis.

Personalization

The Agent maintains a personal memory for each user. Preferences expressed in conversation are stored and applied automatically in all subsequent sessions — without any manual configuration.

Examples:

What the user says	What the Agent remembers
"Always respond in English."	Response language
"I prefer horizontal bar charts."	Default chart orientation
"Keep answers concise."	Response verbosity
"Use a 12-month window for trend analysis."	Default time range

Users can view, update, or clear their stored memory at any time:

• "What do you know about me?"
• "Forget that I prefer English."
• "Clear everything you know about me."

Capabilities by Role

What the AI Agent can do depends on the role of the user in the current session.

Capability	User	Builder	Admin
Generate charts from natural language	✅	✅	✅
Manage own sessions and prompts	✅	✅	✅
Manage charts and dashboards	—	✅	✅
Create and configure datasources	—	✅	✅
Create and manage AI models	—	✅	✅
Update semantic layer descriptions	—	✅	✅
Schedule and manage jobs	—	✅	✅
Manage users and groups	—	—	✅
Monitor platform instance & license	—	—	✅

For a full list of example requests per role, see AI Agent.

Multi-Turn Conversations

The AI Agent maintains the conversation history within a session. Context from earlier messages is carried forward, so users can ask follow-up questions without repeating themselves:

1. "Show me revenue by region." → Chart rendered.
2. "Now filter to Q1." → Agent applies the filter to the same chart.
3. "Which region performed best?" → Agent answers based on the chart data.

Best Practices

• Use a capable Provider Model. Chart generation and semantic layer reasoning require a strong model. Test with your intended model before sharing the Custom Model with users.
• Configure a focused System Prompt. A clear System Prompt helps the Agent stay on topic and improves the quality of chart generation.
• Connect all relevant data sources first. The Agent can only work with data that is connected and has completed semantic extraction.
• Share the model selectively. Because Agent Mode allows write operations (for Builders and Admins), share Agent Mode models with the appropriate user groups.