Agentic AI systems are being adopted by a growing number of organizations to raise productivity and take over repetitive work. As these systems mature and move into production, much of their value comes from the ability to operate autonomously and make sound decisions on a user’s behalf. But wider adoption brings new challenges, two of which stand out: a shortage of skilled talent, since building and managing effective agentic systems demands deep technical expertise, and the difficulty of keeping up with a constantly changing set of frameworks and tools.
Kasal, a Databricks Labs project, was created to address those challenges. It is an agent-driven platform that lets users of varying skill levels design, develop, and deploy agentic AI systems through a visual interface. Non-experts can build sophisticated systems through an intuitive UI, while experts can prototype quickly and export their agents to code for deeper customization later. The stated goal is to make agentic AI accessible to both groups within enterprise environments.
What Kasal is
Kasal is a UI-first framework for designing, running, and observing single- and multi-agent workflows. Instead of hand-writing orchestration code, a user can drag and drop agents onto a visual canvas or describe the goal to a conversational assistant, and Kasal builds the workflow automatically; tools and data can then be connected to it. Under the hood it builds on CrewAI, the open-source Python framework for orchestrating agents, wrapped in a Databricks-friendly application layer that handles authentication, deployment, and monitoring. A broader survey of building blocks for this kind of work appears in this overview of AI agent frameworks.
Why Kasal matters
Kasal brings together three core capabilities: a visual workflow designer driven by agents, deep integration with Databricks, and an extensible toolkit. Visual orchestration makes agents, handoffs, and branching tangible, which makes the logic easier to review with non-developers and to audit afterward. For extensibility, Kasal offers first-class support for MCP servers, Genie spaces, Agent Bricks, custom APIs, and data connectors, and it can export flows as notebooks so that engineers have full transparency into the underlying logic and can extend it freely. It also provides dual-layer observability: business users can follow execution timelines that show workflow status, agent interactions, and task progress, while MLflow tracing lets engineers debug model calls and agent behavior. When deployed as a Databricks App, it uses on-behalf-of authentication and persistence options such as SQLite or Lakebase for user isolation.
Getting started with Kasal
A typical journey begins by describing the desired agent to Kasal — for example, asking it to create a plan that produces a tailored sales pitch deck. Kasal then uses its prompts and language models to generate a structured, often hierarchical plan. In sequential mode, the agents execute one after another in a defined order, while other configurations allow more complex coordination. The most direct way to begin is to install Kasal from the Databricks Marketplace into a workspace, a one-click setup that adds it as a managed Databricks App with built-in governance and automatic updates.
What can be built with Kasal
Users are already building a range of single- and multi-agent systems, including data-analysis pipelines that query, analyze, and visualize data; content-creation systems where agents collaborate on research and writing; business-process automation with workflows that adapt and make decisions; and research-and-development agents that gather, synthesize, and present insights.
Limitations and what to watch
A visual layer lowers the barrier to building agents but does not remove the underlying complexity. Multi-agent workflows can behave unpredictably, so the observability and tracing features matter as much as the design canvas, and outputs should be reviewed before being acted on. Because Kasal is tightly integrated with Databricks, it is best suited to organizations already working in that ecosystem, and as a Databricks Labs project its features and support model may evolve. As with any agentic system that touches enterprise data, access controls, authentication, and governance need to be configured deliberately rather than assumed.
