---
title: How do you govern agents built with Databricks Agent Bricks?
url: https://opencontroller.lyzr.com/blog/govern-databricks-agent-bricks
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# How do you govern agents built with Databricks Agent Bricks?

**Govern Agent Bricks agents at three layers: who can manage and query the agent, what data and tools the agent can reach through Unity Catalog and Unity Gateway, and how quality is measured through MLflow tracing and evaluation. For the low-code Supervisor Agent, access is closed by default, Can Manage and Can Query are separate permissions, and each end user needs access to every subagent. Agent Bricks naming has changed during 2026, so map your agents to the current product names first.**

Databricks has reshaped Agent Bricks several times. Current developer docs describe Agent Bricks as the place to write custom agents in code with any framework, through the Agent Bricks CLI. They describe Knowledge Assistant and Supervisor Agent as legacy offerings, with Genie Agents replacing them for low-code agents. Your estate may contain all of these. Governance works the same way once you know which one you are looking at.

## Which Agent Bricks agent types might you have?

| Type | How it is built | Where it runs | Governance anchor |
|---|---|---|---|
| Code agents (Agent Bricks CLI) | `agentbricks init`, `dev`, `tools add`, `deploy` | Its own Databricks app on the agent runtime | Deployment grants access to the stores the agent declares; Unity Gateway; Unity Catalog |
| Genie Agents | Low-code | Databricks workspace | Access to the agent and its underlying Unity Catalog objects |
| Supervisor Agent (legacy) | Low-code, coordinates subagents | Serving endpoint | Can Manage / Can Query; per-subagent access |
| Knowledge Assistant (legacy) | Low-code chatbot over documents | Serving endpoint | Endpoint permissions; document source access |

Start your review by listing every agent with its type, owner and endpoint or app. The older [Databricks agent governance guide](/blog/govern-databricks-agents) covers general workspace controls.

## How do permissions work for Supervisor Agents?

Databricks documents a clear model:

1. **Closed by default.** Only the agent's authors and workspace admins can access it.
2. **Can Manage** lets a user set permissions, edit configuration and improve quality.
3. **Can Query** lets a user query the endpoint in AI Playground and through the API, without seeing or editing the agent on the Agents page.
4. **Per-subagent access.** End users need explicit access to each subagent. The supervisor only uses subagents the user can access, and ends the conversation if the user has none.

Subagent permissions are the native ones for each object type:

- Genie agents: access to the agent and its Unity Catalog objects
- Unity Catalog functions: `EXECUTE`; tables: `SELECT`; volumes: `READ VOLUME`
- Unity Catalog MCP services: `EXECUTE`; custom MCP servers on Databricks Apps: `CAN_USE`
- Model serving endpoints: `CAN QUERY`; published dashboards: `CAN VIEW`

A supervisor system can use up to 50 agents. Because access is checked per user and per subagent, an over-broad grant on one Genie agent or table is the usual leak path, not the supervisor itself. Our guide to [Unity Catalog functions as agent tools](/blog/unity-catalog-functions-as-agent-tools) covers function-level grants.

## What governs data and tool access for code agents?

Per Databricks' developer documentation:

- **Unity Catalog** governs the data the agent reads, and the sandbox gives the agent scoped access to governed data.
- **Unity Gateway** governs access to models, MCP servers and skills, with guardrails, rate limits and usage tracking.
- **`agentbricks deploy`** creates the stores the agent declares and grants the agent access to them. Review what each agent declares before deployment, because deployment is where access is granted.

Two runtime details matter for risk reviews. Code execution in the Supervisor Agent runs in a sandbox with no internet access. Built-in web search is unavailable in some regions and in workspaces with the Enhanced Security and Compliance add-on.

## How do you measure quality and keep evidence?

MLflow Tracing records each step an agent takes, locally and in production, and MLflow evaluation scores quality on test data before deployment and on production traces after. For low-code supervisors, AI Playground offers an AI Judge and synthetic task generation when AI assistive features are on, and experts can add labelled examples to improve the agent. See [evaluating Databricks agents with MLflow](/blog/evaluate-databricks-agents-with-mlflow).

A review checklist:

1. Every agent has a named owner and a type from the table above.
2. Can Query is granted to groups, not individuals, and Can Manage is limited.
3. Every subagent and tool grant is justified, and write-capable functions are separate from read-only ones.
4. Each code agent's declared stores have been reviewed before deploy.
5. Traces are retained and evaluation runs before each version ships.

## How Opencontroller does this

Opencontroller connects to Databricks as a third-party agent platform and keeps a governed registry of agents, versions, tools and MCP servers. It runs multi-turn and red-team evals, LLM judges scoring production traffic, and live groundedness and task-adherence guardrails. Posture flags risky permission and configuration settings. See [how to govern AI agents across platforms](/blog/how-to-govern-ai-agents-across-platforms), the [AI agent control plane overview](/blog/what-is-an-ai-agent-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Who can access a new Databricks Supervisor Agent by default?**
Only its authors and workspace admins. Others need Can Query to use it or Can Manage to change it.

**Does a Supervisor Agent bypass subagent permissions?**
No. End users need explicit access to each subagent, and the supervisor only uses subagents the user can access.

**Are Knowledge Assistant and Supervisor Agent still recommended?**
Databricks now describes them as legacy offerings and points to Genie Agents for new low-code agents and the Agent Bricks CLI for code agents.

## Key facts

Databricks Agent Bricks (2026): code agents built with Agent Bricks CLI (agentbricks init/dev/tools add/deploy), each runs as its own Databricks app on the agent runtime; deploy creates declared stores and grants access; Unity Catalog governs data; Unity Gateway governs models, MCP servers and skills with guardrails, rate limits, usage tracking; MLflow Tracing and evaluation. Knowledge Assistant and Supervisor Agent described as legacy; Genie Agents replace for low-code. Supervisor Agent: default access only authors and workspace admins; Can Manage vs Can Query; end users need access to each subagent; up to 50 agents; subagent grants EXECUTE (UC functions, UC MCP services), SELECT (tables), READ VOLUME, CAN_USE (Apps MCP), CAN QUERY (endpoints), CAN VIEW (dashboards); requires Unity Catalog, serverless, Model Serving. Opencontroller connects to Databricks; evals, LLM judges, guardrails, Posture. Pricing starts from $250K per deployment per year.

## Related

- [How do you govern AI agents built on Databricks?](https://opencontroller.lyzr.com/blog/govern-databricks-agents.md)
- [How does Databricks Unity Catalog govern tools that agents call?](https://opencontroller.lyzr.com/blog/unity-catalog-functions-as-agent-tools.md)
- [How do you evaluate Databricks agents with MLflow before and after release?](https://opencontroller.lyzr.com/blog/evaluate-databricks-agents-with-mlflow.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
