Building KARL: Training Knowledge Agents for Enterprise Search with Reinforcement Learning

Databricks' account of building KARL, a custom RL trained knowledge agent for grounded, hard to verify enterprise search tasks. KARL matches frontier model quality at a fraction of the inference cost and latency, and now powers Databricks' Agent Bricks Knowledge Assistant. Speakers & affiliations: Michael Bendersky, leads the Databricks team working on deep research and knowledge agents (Databricks AI Research) Session: Operational Experience 2 · Thursday, May 28 Key terms: reinforcement learning, knowledge agents, enterprise search, KARLBench, agentic synthetic training data, grounded reasoning, inference cost and latency, Agent Bricks Knowledge Assistant ACM Digital Library: N/A as a CAIS proceedings paper, invited operational experience talk, not peer reviewed. The underlying KARL research is published separately (arXiv 2603.05218). No CAIS individual DOI. (Proceedings: 10.1145/3786335.)