The hardest thing to teach a search system is restraint. Most AI agents, when tasked with finding information, keep pulling context until they hit a token limit or a timeout. Databricks decided to train one that stops when it’s actually done.
The result is KARL, a retrieval-augmented generation agent built on reinforcement learning that matches the performance of Claude Opus 4.6 while costing 33% less and running with 47% lower latency.
Teaching an agent to stop
The core insight behind KARL is that search inefficiency is usually a timing problem. Traditional RAG systems retrieve more context than they need, process redundant information, and burn compute doing it. KARL uses reinforcement learning to recognize when additional retrieval stops adding value, then halts the search instead of continuing by default.
Databricks pairs this with context compression, a technique that lets the agent condense what it has already retrieved before deciding whether to fetch more.
What Agent Bricks adds to the picture
KARL sits inside a larger framework Databricks introduced called Agent Bricks, which launched in September 2026 as a platform for building auto-optimized, domain-specific agents. The premise is that enterprises should not need to hand-tune every model they deploy. Agent Bricks handles that optimization automatically, adjusting agent behavior to fit specific task types defined in plain language.
Since the launch, more than 100,000 agents have been built on the platform. The framework supports major model providers including Claude and GPT variants, with governance routed through Databricks’ Unity Catalog.
Databricks has also layered in a reranking capability within its AI Search product. Reranking is the process of taking an initial set of retrieved documents and re-scoring them for relevance before passing them to the language model. The accuracy gains on enterprise benchmarks have averaged 15 percentage points.
Accuracy at the frontier without frontier prices
The benchmark numbers Databricks is citing internally show accuracy improvements from roughly 32% to over 90% on certain task types when using specialized agent techniques like parallel thinking and multi-LLM designs.
Matching Claude Opus 4.6 on retrieval and reasoning tasks while undercutting its cost by a third positions Databricks in an interesting place in the enterprise market. The company is not selling a foundation model. It is selling the infrastructure that makes foundation models usable at scale.