Databricks
Databricks AI agents: how they run and what a team should govern
A plain guide for data and platform leads who are considering an agent on Databricks. Every fact about the platform comes from Databricks' own documentation as read on 11 November 2026. Features change, so check the cited page before you act.
A Databricks AI agent is an AI agent that a team builds and runs on Databricks, where Agent Bricks is what Databricks calls its agent developer platform. Databricks' documentation, last updated on 13 October 2026, says you run an agent with an agent server on managed compute, send requests through the invocation API, the Responses API or SQL, record each step with tracing, and govern access to models, MCP servers and skills with guardrails, rate limits and usage tracking. Our view: before a first agent goes live, a team should decide who owns it, what it may reach and who reviews its traces. Source read 11 November 2026.
plain answer
Databricks AI agent: what is it?
In Databricks' documentation, "Agent Bricks is the Databricks agent developer platform." A Databricks AI agent is therefore an agent you build and run there. The page does not define what an AI agent is, so we do not offer a Databricks definition.
The page lists parts of the platform: an agent runtime that hosts agents built with any framework or harness, stateful or stateless; memory so an agent can recall what it learned in earlier conversations; and a sandbox, described as an isolated environment whose home directory persists across sessions.
Source: Databricks, Develop agents on Databricks (https://docs.databricks.com/aws/en/generative-ai/agent-framework/build-genai-apps), page label last updated on Oct 13, 2026, read 11 November 2026.
how to
How do you run and call a Databricks AI agent?
Databricks says you run your agent with an agent server on managed compute, and that AgentKit adds durable execution so that runs survive restarts and crashes. You send requests to deployed agents through the invocation API, the Responses API, or SQL.
For quality, the page lists tracing, which records each step an agent takes so you can debug and improve it, and evaluation, which measures agent quality with LLM judges, custom scorers and expert feedback. It also says you can store and govern traces at scale in Unity Catalog.
Source: Databricks, Develop agents on Databricks (https://docs.databricks.com/aws/en/generative-ai/agent-framework/build-genai-apps), read 11 November 2026.
security
What should a team govern before a Databricks AI agent goes live?
Databricks lists a governance feature: govern access to models, MCP servers and skills, with guardrails, rate limits and usage tracking. That gives you controls. It does not decide for you who may use the agent or what it may do.
Our view: settle four things first. Who owns the agent and answers when it is wrong. Which data and tools it may reach, kept to the smallest set the job needs. Which actions need a person's approval before they happen. Who reads the traces, how often, and what they do when a trace looks wrong.
Sources: Databricks, Develop agents on Databricks (https://docs.databricks.com/aws/en/generative-ai/agent-framework/build-genai-apps), read 11 November 2026. The four decisions are Kriv's view.
how to
How should a first agent be scoped?
Our view: pick one workflow that staff already document well, with one owner and a clear review step. Run it with a person checking every output, read the traces for a few weeks, and widen access only after the results justify it.
Compare what the agent does with what a simpler, fixed workflow would do. If a fixed workflow does the job, an agent may not be needed.
kriv role
Where Kriv helps
Kriv AI helps data teams scope an agent on Databricks: the workflow, the access it needs, where a person approves, how traces are reviewed and who owns it afterwards. Engagements are scoped and priced after a short discovery call.
Kriv did not build Agent Bricks and does not replace Databricks' documentation. Kriv has no public case study for an agentic build, so this page describes a method, not a delivered result.
limits
What this page could not confirm
We read one Databricks documentation page, labelled for Databricks on AWS. We did not read the pricing, regional availability, preview or release status of individual Agent Bricks features, or the pages for other clouds, and we did not test an agent. The page does not mention Mosaic AI Agent Framework by name. Check Databricks' current documentation for anything not cited above.
Related resources
Continue exploring
Straight answers
Frequently asked questions about Databricks AI agents: how they run and what a team should govern
What is Agent Bricks?
Databricks' documentation says Agent Bricks is the Databricks agent developer platform (page label last updated on Oct 13, 2026, read 11 November 2026).
How do you call a deployed Databricks agent?
Databricks says you send requests through the invocation API, the Responses API, or SQL (read 11 November 2026).
Can you see what a Databricks agent did?
Databricks lists tracing, which records each step an agent takes, and says you can store and govern traces at scale in Unity Catalog (read 11 November 2026).
Does Databricks control what an agent can access?
Databricks lists a feature to govern access to models, MCP servers and skills, with guardrails, rate limits and usage tracking (read 11 November 2026). Deciding who may use an agent and what it may do is still the team's job.
Can Kriv help us plan a first agent on Databricks?
Kriv AI scopes the workflow, the access, the approval steps and the trace review. Scope and price are set after a short discovery call.
Talk to the team that would do the work
Bring your requirements to a working session with the person who'll actually deliver.
Book a Discovery Call