Photo by RDNE Stock project on Pexels
Databricks Agent Bricks (Custom Agents, Document Intelligence and Supervisor Agent GA)
Databricks' enterprise agent platform now has its headline pieces generally available, and its Unity Catalog-rooted governance story is the most convincing part of the pitch. Without published pricing or performance benchmarks, though, this is a promising platform we can't yet fully grade.
Pros
- + Governance is the core design, not an add-on: Unity Catalog tracks every agent, model, tool and MCP server with clear ownership and granular permissions
- + On-Behalf-Of authentication validates each data fetch and tool call against the human user's real permissions, avoiding over-privileged service accounts
- + Multi-vendor model access (OpenAI, Anthropic, Google, Meta, open source) with routing and automatic fallbacks means no re-architecting to switch models
- + Supervisor Agent sensibly separates structured data, unstructured data and tool use instead of forcing one retrieval pattern
- + Per-user and per-team rate limits give finance and platform teams a real throttle
- + Open developer path via familiar IDEs and frameworks like LangChain and LangGraph
- + Built-in evaluation and SME feedback loops treat agent quality as a first-class concern
- + Omnigent's ability to compose Claude Code, Codex and custom agents above existing harnesses is a genuinely differentiated approach
Cons
- − No pricing published at all, which is awkward for a platform that markets cost control
- − Adoption figures (100,000+ agents, 1 quadrillion tokens/year) are vendor-reported scale metrics, not evidence of accuracy or reliability
- − No accuracy, latency or benchmark data for Document Intelligence or the supervisor routing
- − Value depends heavily on already being invested in Unity Catalog and the Databricks platform — limited appeal outside that ecosystem
- − Several newer capabilities, including AI Gateway and the Omnigent workflow, are described at a high level with little operational detail
- − Breadth of the stack (supervisor, Genie Spaces, Knowledge Assistants, MCP servers, gateway, MLflow) implies real learning curve and lock-in
Agent Bricks is Databricks' attempt to answer the question every enterprise AI team is currently stuck on: not "can we build an agent?" but "can we let one loose on our actual business data without creating an audit nightmare?" With Custom Agents, Document Intelligence and the Supervisor Agent all reaching general availability — alongside a new AI Gateway — the platform has moved from interesting demo to something Databricks is clearly asking customers to standardise on.
The shape of the thing is sensible. A Supervisor Agent sits on top as a managed orchestration layer, reading a user's question and routing it between Genie Spaces for structured data, Knowledge Assistant agents for unstructured content, and MCP servers for tooling. Underneath, Unity Catalog does the unglamorous work: granular permissions, per-user and per-team rate limits, and an inventory of every agent, model, tool and MCP server with an actual owner attached. On-Behalf-Of authentication means the supervisor acts as a transparent proxy for the human, so each data fetch is checked against that person's existing permissions rather than a service account's.
This is squarely a platform for existing Databricks shops — data platform teams, ML engineers and the compliance people who have to sign off on their work. If your governance model already lives in Unity Catalog, the value proposition is obvious. If it doesn't, much of what makes this compelling simply isn't available to you.
Databricks' own framing is the most honest thing about this launch: the agent loop is roughly 1% of the work, and the real problems are Choice, Context and Control. Agent Bricks is built around that thesis rather than around a flashy agent builder, and it mostly delivers on it. Multi-provider model access — OpenAI, Anthropic, Google, Meta and open source through one interface, with intelligent routing and automatic fallbacks — is the right hedge in a market where the best model changes every few months, and being able to switch without re-architecting is worth real money. Developer ergonomics look respectable too, with support for familiar IDEs and frameworks like LangChain and LangGraph, plus built-in evaluation and subject-matter-expert feedback loops rather than treating quality as someone else's problem.
The Omnigent story is the most forward-looking piece here: running agents above existing harnesses so Claude Code, Codex and custom agents can be composed into one workflow, with policy enforced at runtime through the Unity AI Gateway and sessions traced in MLflow. That is a genuinely unusual position — meta-orchestration rather than yet another walled garden.
Where I have to hold back is evidence. Databricks cites thousands of organisations running production agents and names serious customers — Workday, Virgin Atlantic, Zapier, EchoStar, AstraZeneca, 7-Eleven, Fox Corporation and Block — along with over 100,000 agents built and more than a quadrillion tokens processed annually. Those are impressive logos and impressive numbers, but they are vendor-supplied adoption metrics, not quality metrics. Nothing here tells me how accurate Document Intelligence is on messy real-world documents, how well the dynamic supervisor pattern holds up on ambiguous multi-hop questions, or what latency looks like once routing and fallbacks are in play. Most conspicuously, there is no published pricing at all, which for a platform whose own pitch includes cost control is a gap buyers will feel immediately.
The verdict, then, is a qualified one. The architecture is thoughtful, the governance model is the best-articulated part and the GA milestones mean this is no longer a bet on a roadmap. But I can't rate the execution of something I can only see described. Four stars for a well-designed, well-positioned platform with a clear point of view — withheld marks for the absence of pricing and any independent measure of how well it actually performs.