Databricks introduces KARL AI agent to optimize search efficiency

Databricks has developed KARL, an AI agent designed to improve search efficiency by recognizing when it has gathered sufficient information to stop retrieving data. Unlike traditional retrieval-augmented generation (RAG) systems that often continue searching until hitting token limits or timeouts, KARL utilizes reinforcement learning to determine when additional context no longer adds value.

Technical benchmarks indicate that KARL matches the performance of Claude Opus 4.6 while operating at 33% lower cost and 47% lower latency. The agent employs context compression to condense retrieved information before deciding whether further searches are necessary.

KARL is integrated into Agent Bricks, a platform launched in September 2026 for building auto-optimized, domain-specific agents. The platform allows enterprises to deploy agents that adjust their behavior based on plain-language task definitions without manual fine-tuning. Since its launch, more than 100,000 agents have been built using the Agent Bricks framework, which supports major models including GPT variants and Claude, with governance managed through the Unity Catalog.