The promise of multi-agent systems often hits a wall when it comes to presenting complex data. Agents typically churn out tables, a format ill-suited for quick comprehension on platforms like Microsoft Teams. Databricks is tackling this head-on, integrating Vega-Lite with its Agent Bricks and Unity Catalog Functions to enable visualizations that are both portable and governed.
This new capability allows AI agents to generate and refine charts directly, transforming raw data into actionable visual insights. It moves beyond the limitations of simple text tables, enabling richer communication where stakeholders actually work.
Bridging the Visualization Gap
The core challenge lies in delivering insights consistently across diverse platforms. Each external system has its own visual language, making direct integration difficult. Databricks' agent framework, extensible via Unity Catalog Functions and the Model Context Protocol (MCP), allows developers to overcome these limitations.
The system comprises Supervisor Agents that orchestrate specialized tools. These include Genie Spaces for SQL queries, Knowledge Assistant agents for document analysis, Unity Catalog Functions for custom logic, and MCP servers for third-party integrations. This architecture excels at decomposing complex requests.
Governed Visualizations for Agents
Databricks Agent Bricks, a key component for production-ready AI, facilitates this process. These agents can now leverage Vega-Lite, a declarative JSON specification for creating charts. This allows agents to produce visualizations as easily as they output text or data.
Using Vega-Lite offers several advantages: it's API-native and renders consistently across applications, it's LLM-friendly due to its compact nature, and its schema-based validation aids in error correction. Crucially, declarative JSON avoids the security risks associated with generated plotting code.
Unity Catalog Functions act as a secure, reusable layer, centralizing visualization logic. Agents can call these functions, which validate input data, infer schemas, select chart types, and construct the Vega-Lite JSON specification.
Real-World Impact
Teams in finance, healthcare, and sales are already exploring these capabilities. A common scenario involves a CFO asking for Q4 revenue performance against forecasts directly in Microsoft Teams.
The Supervisor Agent decomposes this request, routing parts to Genie agents for data retrieval and a Unity Catalog Function for visualization generation. The result is a cohesive response within Teams, including a text summary and interactive Vega-Lite charts. This bypasses the need for manual data export and charting, drastically reducing the time to insight.
Early pilot results are compelling. Users are seeing up to 90% faster time to insight, 3-4 times more questions answered per session, and approximately double the adoption rate among non-technical users. Agent response satisfaction also jumped nearly 40%.
As multi-agent systems become integral to enterprise workflows, the ability to not just compute answers but to visually present them will be paramount. This move towards visual intelligence in multi-agent systems visualizations is a significant step forward.
