Databricks KARL trains enterprise agents with RL

KARL is a notable shift away from brittle, hand-coded agent workflows toward training retrieval behavior directly, which is exactly where enterprise AI needs to go. If the reported results hold up, this is less a chatbot story than a blueprint for how companies will build production knowledge agents.

• –KARLBench matters almost as much as the model itself because it packages six messy enterprise search and synthesis regimes into a benchmark that reflects real knowledge work better than narrow QA evals
• –The paper’s strongest claim is multi-task generalization: training across heterogeneous search behaviors appears to outperform single-benchmark optimization and helps on out-of-distribution tasks
• –Databricks is betting that synthetic, tool-using data generation plus off-policy RL can replace a lot of prompt-and-workflow engineering, which could lower the maintenance burden for enterprise agent systems
• –The official launch framing also matters: Databricks is positioning KARL as a preview tied to custom RL, suggesting this research is intended to become part of the platform rather than remain a one-off lab demo