Agentic Reasoning in Practice: Making Sense of Structured and Unstructured Data

Databricks Agent Bricks Supervisor Agent demonstrates multi-step agentic reasoning across structured and unstructured enterprise data for complex knowledge-intensive tasks.

• •Supervisor Agent achieves 20%+ improvement over baselines on STaRK and KARLBench benchmarks through declarative agent building and multi-step reasoning.
• •The agent decomposes complex queries into sub-tasks, routing to appropriate tools: Genie for structured data retrieval and Knowledge Assistant for unstructured document search.
• •Superior performance compared to single-turn vector similarity search with LLM-based reranking, with largest gains on exhaustive analysis and self-correction tasks (+38% biomedical, +78% entity search).
• •Built on aroll framework, SA handles thousands of concurrent conversations, enables tool and instruction configuration without custom code, and incorporates advanced orchestration and context management.