Context drives agent accuracy
Databricks uses MosaicML technology to provide the business context required for AI agents. The company acquired MosaicML for $1.3 billion to improve neural network training and inference. This integration allows enterprises to build custom models on proprietary data. MosaicML built the Composer tool to speed up the training of neural networks for standard AI programs like GPT. The company also released the MPT-30B language model, which has 30 billion parameters. MosaicML focuses on building experts through domain-specific models rather than generalists. Databricks testing showed that applying ontology context increased answer accuracy from 52.4% to 84.5% for an internal 28-question enterprise suite. You already know that models fail without reliable business context. The system learns which definitions teams rely on to ensure agents use the same logic as human workers.
Real-time data and governance
LTAP combines transactional and analytical workloads on a single data copy using Lakebase and Unity Catalog. This architecture prevents the need for separate ETL pipelines or replicas by bringing OLTP and OLAP together. By using the Reyden engine, Lakehouse//RT delivers millisecond latency for agentic workloads.
| Feature | Specification/Detail |
|---|---|
| Lakehouse//RT Latency | 10ms to sub-100ms |
| Lakehouse//RT Throughput | 12,000 queries per second |
| Agent Bricks Usage | 100,000+ agents |
| Agent Bricks Volume | 1 quadrillion tokens per year |
| Genie transition date | July 6, 2026 |
The Unity AI Gateway provides runtime control for every model, agent, and tool call by allowing administrators to set hard spend caps and use smart routing to select models based on cost or performance for every request. This gateway includes guardrails for PII exposure and prompt injection. Unified governance through Unity Catalog provides a four-level namespace where every asset has a single cross-cloud address under one set of policies and one audit trail. These definitions ensure agents and human users see the same metrics across SQL, BI, and APIs. Lakeflow provides a unified ingestion and orchestration layer to consolidate specialized tools into one platform.
Managing agentic costs
Genie ZeroOps monitors data and AI assets to detect pipeline failures. It uses Unity Catalog lineage to perform root cause analysis and proposes fixes in a sandbox for human approval. Genie transitions to a pay-as-you-go model on July 6, 2026. Admins set budgets and alerts via the Unity AI Gateway. These autonomous workflows introduce new operational burdens. Many companies struggle with the cost of model inferencing and the need for custom engineering to connect modern model endpoints with legacy databases. How will organizations handle the rising complexity of managing thousands of autonomous agents?
Agent Bricks facilitates the deployment of agents on harnesses like LangGraph, CrewAI, and Claude Code SDK. Large organizations like AstraZeneca and 7-Eleven have already built over 100,000 agents using the platform. The platform handles the technical debt of token capacity, deployment, security, evaluation, monitoring, and context. Genie One is a data-smart AI coworker that works inside or outside Databricks. It connects to tools like Slack, Teams, and Gmail. Users can use Genie App Builder to create internal apps through vibe-coding. CustomerLake provides an agentic Customer Data Platform in private preview that uses Profile Agents and Campaign Agents to build Customer 360 profiles. These agents use Agentic Identity Resolution to blend deterministic and probabilistic workflows. This tool uses data that never moves to build audiences and recommend next-best actions.
