We recently introduced Dashboard Agents ( Preview) in Looker, a new feature that allows viewers to ask questions directly within a dashboard and receive immediate, context-aware answers about their data.
To help users get the most accurate answers, data analysts can configure these agents to match specific business needs. This post explains how Dashboard Agents work behind the scenes and shares best practices for setting them up for high accuracy, optimal performance, and a great user experience.
Some best practices to get started with Dashboard Agents configurations
1. Optimize the Dashboard Layout
A dashboard provides the core context an agent uses to understand your business data and answer user questions. A clean, focused dashboard directly improves the agent’s performance. more details on building a performant dashboard visit our help center article
-
Manage Tile Density: Aim max 25 tiles per dashboard. Too many tiles can increase token consumption and processing time, which leads to slower response times for your users.
-
Preconfigure Filters: Set default filters to give the agent a refined dataset from the start. This prevents the agent from needing extra steps to narrow down data, making the user experience much smoother.
2. Prioritize Descriptive Metadata
Metadata serves as the map the agent uses to navigate your visualizations. Without clear labels, the agent may struggle to find the right data points.
-
Use Unique, Concise Titles: Dashboard agents reference tile titles when citing sources in their answers. Ensure every title is unique and descriptive so the agent can accurately credit the correct visualization.
-
Leverage Subtexts and Labels: Use description fields to provide extra context. This helps the agent understand the purpose of a visualization and exactly what its metrics represent. As a bonus, this also makes your dashboards easier for human audiences to read.
A Sample dashboard with clear titles, filters with default values, labeled axis and field descriptions.
3. Ground the Agent in LookML
Instead of building complex business logic directly into the agent, use a semantic model to house your permanent business rules.
- Centralize Business Logic: Define glossaries and metrics within LookML to establish a single source of truth. This keeps data definitions consistent across your organization and allows you to reuse the same configuration for multiple agents.
Improve data context for self-service dashboards by defining clear field names and detailed descriptions.
By organizing your dashboard canvas, metadata, and semantic models, you create a reliable technical foundation. This ensures the agent delivers precise business intelligence rather than educated guesses.
4. Refine Agent Behavior with Instructions
You can fine-tune how the agent interacts with users by providing tailored instructions in the final configuration layer.
-
Identify the Intended Audience: Tell the agent who will be using the dashboard so it can adapt its response structure to their needs.
-
Establish Tone and Formatting: Specify how you want answers structured or styled to match your organization’s preferences.
-
Define Business Terms and Defaults: Use instructions to clarify acronyms or set fallback rules. For example, you can tell the agent how to handle ambiguous timeframes.
-
Example: “Interpret 'Rev’ as ‘Revenue’.”
-
Example: “If no timeframe is mentioned, default to the current quarter.”
-
-
Prioritize Clarity: Keep instructions concise and free of conflicting rules to prevent the agent from becoming confused.
-
Use Structured Formats: Write your instructions in Markdown or YAML. AI models process structured text much more effectively than dense paragraphs.
To achieve optimal Dashboard agent performance, it is crucial to iterate on your instructions and thoroughly preview the agent’s behavior before saving your final configurations. “Preview” in Manage agent enables you to assess agent responses before saving your final configurations
Example: Agent Instructions for Cymbal Pets Business Pulse
The following block contains sample instructions tailored for the Cymbal Pets Business Pulse dashboard - a business performance overview dashboard.
**Intended Audience:**
- Sales Executives and Store Managers
**Tone and Style:**
- Maintain an analytical and professional tone in all responses.
**Response Structure:**
- Provide a sentence for key takeaways
- Surface data deepdives
- Provide business insights
- Recommend next steps for analysis
**Business Logic & Terminology:**
- Interpret "Rev" as "Revenue"
- Interpret "YTD" as "Year to Date”
**Default Behaviors:**
- if no date filter is explicitly mentioned in user prompt default to current calendar year.
**Metric Breakdown:**
- If no specific dimensions are mentioned in user prompt Prioritize and highlight performance metrics breakdown for the following core categories: Food, Accessories, and Health & Wellness.
Carefully configuring your dashboard layout, descriptive metadata, semantic models and leveraging Agent instructions to customize your agent ensures the agent delivers tailored, precise and reliable data insights.
Looker Admins can enable the dashboard agent capability by toggling “Enable Chat with Dashboard” within the Gemini in Looker settings. Once enabled, users with right permissions will see the Manage Agent icon in the Dashboard Agents Side Panel to begin configuring the agents. See our help center for more information.