Databricks launches Agent Bricks with 100k agents

Databricks is accelerating the transition from conversational chatbots to action-oriented AI agents through its Agent Bricks platform. According to the company, organizations have already deployed more than 100,000 agents on the service, which collectively process over one quadrillion tokens annually. This rapid adoption aligns with projections from Gartner, which estimates that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, a massive jump from less than 5% in 2025. At the core of this shift is tool calling, the structural mechanism that allows models to dynamically select and execute external APIs, databases, and code.

For developers and data practitioners, tool calling represents a shift away from rigid, rule-based integrations. Instead of writing brittle point-to-point connections, developers can use the Python-based Agent SDK to let models evaluate user intent and select the appropriate tool. For instance, an agent might process a query about Q2 sales by pulling raw data, calculating that revenue reached $14.3 million—representing a 12% increase over Q1—and emailing the formatted report. To evaluate the reliability of these models, practitioners rely on benchmarks like the Berkeley Function Calling Leaderboard (BFCL) across various LLM providers.

Agent Bricks simplifies this orchestration by natively supporting the Model Context Protocol (MCP), an open standard for tool integration, alongside Unity Catalog for strict data governance. The platform integrates with popular developer frameworks, including LangGraph, CrewAI, Agno, the Claude Code SDK, and OpenAI Agent SDKs. By managing permissions, prompt injection prevention, and sensitive data filtering within a single control plane, the system allows teams to safely transition agents from passive text generators into active enterprise workers.