How do I run an agent fleet with centralized controls?

Summary

• One control plane for many agents. On Databricks, you run a fleet of AI agents from a single governed control plane built on Agent Bricks, the Mosaic AI Agent Framework, and the Unity AI Gateway. See Agent Bricks.
• Unity Catalog is the unified registry. Every model, agent, MCP server, and tool is registered and governed as a Unity Catalog securable, with the same access, lineage, and audit controls you use for data. See Governing AI agents at scale with Unity Catalog.
• Central traffic controls. The Unity AI Gateway routes every model and tool call through one control plane and enforces rate limits, budgets and spend caps, usage tracking, and smart routing across models. See how AI governance works.
• Coordinate agents with a supervisor. The Supervisor Agent orchestrates up to 20 sub-agents; the invoking user's identity propagates on-behalf-of through every sub-agent down to the deepest tool call.
• Fleet-wide observability. MLflow tracing automatically logs each agent's full execution graph, tool calls, latency, and token usage, and every action is captured in audit logs for end-to-end monitoring.

How do I run an agent fleet with centralized controls?

Running many AI agents in production means controlling them from one place rather than managing each in isolation. On Databricks, you operate an agent fleet from a single governed control plane that combines Agent Bricks, the Mosaic AI Agent Framework, and the Unity AI Gateway. Deployment, orchestration, tools, models, and governance live together, so you keep enterprise control while giving builders the freedom to use any model or framework.

Why Databricks Agent Bricks for a governed agent fleet

• Unity Catalog as the unified registry. Databricks-hosted and external models, custom and external agents, MCP (Model Context Protocol) servers, and enterprise tools are all registered and governed as Unity Catalog securables. You apply the same familiar controls for access, discovery, lineage, audit, and monitoring across every asset, instead of managing them in separate tools. See Governing AI agents at scale with Unity Catalog.
• The Unity AI Gateway control plane. The AI Gateway routes every model-service and MCP-service request from one central control point. It enforces rate limits (QPM and TPM caps per principal), budgets and hard spend caps, usage tracking and cost attribution, and smart routing that selects the right model to balance quality and cost. Traffic splitting and fallbacks distribute requests across model destinations and add failover for availability. See how AI governance works on Databricks and Unity AI Gateway is generally available.
• Service policies for behavior governance. Beyond static grants, service policies evaluate each request and response and return allow, deny, or require-approval based on who is calling and what the request contains.
• Supervisor orchestration for multi-agent systems. The Supervisor Agent provides a managed foundation for orchestrating up to 20 sub-agents. It routes intent to the right sub-agent, governs access through Unity Catalog, and drives continuous improvement from one control plane. The invoking user's identity propagates on-behalf-of through every sub-agent down to the deepest tool call, so row filters, row-level security, and function authorization all see the real user. See the Supervisor Agent.
• Tools governed as securables. Managed and custom MCP servers and function tools are governed in Unity Catalog with fine-grained access control, rate limits, and policies. Attribute-based access control (ABAC) grants let you control tool access by tag, and each sub-agent's tools carry their own grants.
• Fleet-wide observability with MLflow. Every agent invocation is automatically traced to MLflow using OpenTelemetry standards, capturing the full execution graph, reasoning steps, tool calls, retrieved documents, latency, and token usage. Traces are stored in governed tables in your lakehouse for debugging, evaluation, and production monitoring. See Agent Bricks at Data + AI Summit 2026.

Getting started

• Read Governing AI agents at scale with Unity Catalog to see how permissions, policies, and the registry fit together.
• Set up AI governance in the Unity AI Gateway for rate limits, budgets, usage tracking, and routing.
• Build multi-agent coordination with the Supervisor Agent.
• Explore the Agent Bricks platform for building, deploying, and governing agents.

FAQs

What gives you centralized control over an agent fleet?

The Unity AI Gateway is the central control plane: it routes every model and tool call and enforces rate limits, budgets and spend caps, usage tracking, and smart routing, while Unity Catalog governs every agent, model, and tool as a securable.

How do many agents work together?

A Supervisor Agent orchestrates up to 20 sub-agents, routing intent to the right one and governing access through Unity Catalog, with the user's identity flowing on-behalf-of through every sub-agent to the deepest tool call.

How do you monitor agents in production?

MLflow tracing automatically records each agent's full execution graph, tool calls, latency, and token usage into governed lakehouse tables, and Unity Catalog audit logs capture every action for end-to-end observability.