- BFSIBanking, Financial Services & Insurance
- IT & ITESIT services, BPM, GCCs & SaaS
- Retail & e-commerceStore chains, D2C brands & marketplaces
- HealthcareComing soon
- ManufacturingComing soon
Databricks Agent Bricks
The agent demo went well. Then risk asked how you know its answers are right — Databricks Agent Bricks builds, evaluates, serves and governs AI agents inside your Databricks workspace, with hosted Claude, GPT, Gemini and open models — billed per DBU and governed by Unity Catalog.
Buy through TechBag
Same software. Better outcome — at a lower cost.
How it’s rated
Full scoreboard ↓Quick answer
The Databricks platform family
This page covers Databricks Agent Bricks — with Model Serving, Foundation Model APIs, AI Search, managed MLflow and AI Runtime folded in. The rest:
Quick facts
30-second orientation- Product
- Databricks’ umbrella for building, evaluating, serving and governing AI agents and models
- Maker
- Databricks, Inc., San Francisco; its AI products passed a $1.4B revenue run-rate by February 2026
- Folds in
- Model Serving, Foundation Model APIs, AI Search, Agent Framework, managed MLflow and AI Runtime
- Models
- Hosted Claude, GPT, Gemini, Llama, Qwen, DeepSeek, GLM and Kimi, plus your own; the list changes monthly
- Price
- DBUs at each product’s rate; serverless inference is $0.07 in US East, $0.074 in AWS Mumbai and Azure Pune
- Closed
- Knowledge Assistant and Supervisor Agent take no new customers; Databricks points them to Genie Agents
- Deployment
- Serverless, inside a Databricks workspace on AWS, Azure or Google Cloud; no on-premises edition
- Analysts
- Gartner MQ for AI Platforms for Data Science and ML, 2026: Leader, highest on execution, furthest on vision
- India
- Model Serving, provisioned throughput and AI Search run in Mumbai and Pune; fine-tuning is not offered there
- In India via
- TechBag — scoping the use case, a governed pilot, and the quote itemised in INR with GST
Understand AI agent platforms before you buy one
Most product pages skip this. We start here — so you buy a capability, not a buzzword.
What is Databricks Agent Bricks?
Agent Bricks is Databricks’ AI umbrella: build, evaluate, serve and govern agents and models.
Agents stitched from separate AI services vs Databricks Agent Bricks — the honest table
What consolidation actually replaces, dimension by dimension.
| Dimension | Agents stitched from separate AI services | Databricks Agent Bricks |
|---|---|---|
| Models | An API key per model provider | Hosted models behind one serving API |
| Retrieval | A separate vector database to sync | AI Search in the same workspace |
| Quality | Spot checks in a spreadsheet | LLM judges and traces before release |
| Access | Credentials copied into each app | Unity Catalog permissions on every call |
| Billing | A bill from every AI vendor | DBUs on the Databricks invoice |
| What it is NOT | — | A no-code assistant for business users; that is Genie |
The cheapest test is one agent on one governed dataset: build it, judge it on fifty real questions, and read the DBUs it used.
The four pieces of the platform
Vendors love diagrams; buyers need to know what they’re actually operating. Here’s the whole platform, demystified.
Models
Hosted Claude, GPT, Gemini, Llama and open-weight models by pay-per-token, provisioned throughput or batch, plus external providers routed through Unity Gateway.
Agents
Write agents in LangChain, LangGraph, LlamaIndex or plain Python, add retrieval from AI Search and managed memory (Beta), then deploy each one as a serving endpoint.
Evaluation
LLM judges score answers against your criteria and MLflow traces every step; since July 2026 those traces are stored in Unity Catalog beside the data.
Serving
Serverless CPU and GPU endpoints serve agents and custom models, AI Runtime adds serverless GPUs for training (Public Preview), and Unity Catalog governs access.
Hosted models and AI Search feed agents built in open frameworks — LLM judges and MLflow traces prove them, and Model Serving runs them under Unity Catalog.
Nine capabilities. Build, govern, serve.
Databricks Agent Bricks takes an agent from framework code to judged, traced and governed production — next to the data it uses.
Bring LangGraph, LangChain or LlamaIndex
Agents are written in the open-source framework your team already uses, then logged, versioned and deployed through managed MLflow.
A wide hosted model menu
Claude, GPT, Gemini, Llama, Qwen, DeepSeek, GLM and Kimi models answer behind one API, and Databricks revises the list most months.
AI Search for grounding
AI Search, formerly Vector Search, retrieves by vectors or by full text with no embeddings needed; a Standard unit lists at $0.28 an hour.
LLM judges for agent quality
Agent Evaluation scores responses with LLM-as-judge against your own criteria, so a new prompt or model is tested before it ships.
MLflow traces in Unity Catalog
Managed MLflow records each step an agent takes; since 16 July 2026 the traces land in Unity Catalog, governed like any table.
Unity Gateway over models and tools
Unity Gateway decides which models, tools and MCP servers agents may call, tracks spend and applies guardrails; GA since August 2026.
Serverless CPU and GPU serving
Model Serving runs agents and custom models on serverless CPUs at 4 DBU an hour, or on GPUs from one T4 up to eight H100s.
Provisioned throughput, in India
Provisioned-throughput endpoints keep inference inside the India Geo; pay-per-token calls from India need cross-Geo processing switched on.
Serverless GPUs for training
AI Runtime, in Public Preview, rents serverless GPUs for custom training: an A10 lists at $2.50 an hour in US East, $3.27 in AWS Mumbai.
Watch Databricks Agent Bricks in action
Three videos from Databricks’ official channel: a Data + AI Summit 2026 demo, a session on production-ready agents, and a short January 2026 intro that predates the March umbrella.
How Agent Bricks gives developers choice, context and control (with demo)
A 26-minute session with demo, from June 2026, on how Agent Bricks gives developers choice, context and control.
Building Enterprise-Ready Agents using Agent Bricks
A 29-minute session from May 2026 on building enterprise-ready agents with Agent Bricks.
Introduction to Agent Bricks
A two-minute intro from January 2026; it predates the March umbrella and may show agent types now closed to new buyers.
Want a live, India-context walkthrough for your environment?
Book a guided demo →Most agent pilots stall at the risk review. Databricks Agent Bricks judges and traces every answer.
Here’s what genuinely sets it apart — and exactly where it stops.
Agents built next to the governed data
Most agent platforms sit outside the data platform and need data or credentials copied into them. Agent Bricks runs in the same workspace as your tables, so retrieval, tool calls and traces inherit Unity Catalog permissions and lineage, and a reviewer can follow an answer back to the sources it drew on.
Many models, one invoice
Claude, GPT, Gemini, Llama and open-weight models such as Qwen and DeepSeek sit behind one serving API and bill as DBUs on the Databricks invoice. Unity Gateway routes and budgets calls across providers, so switching a model for cost or quality is a configuration change rather than a new vendor onboarding.
Evaluation built in, not bolted on
LLM judges, agent evaluation and MLflow traces stored in Unity Catalog ship in the same product as serving, so quality gates run before release. IDC named Databricks a Leader in its 2025 MarketScape for GenAI evaluation products, and Gartner’s 2026 AI Platforms for DSML Magic Quadrant rated it a Leader.
Where it stops
Everything is consumption-billed in DBUs across several SKUs, so a monthly figure needs a pilot. Knowledge Assistant and Supervisor Agent are closed to new customers. In India, pay-per-token models need cross-Geo processing, Foundation Model fine-tuning is not offered, and AI Runtime is still in Public Preview.
The numbers behind the platform
What your Databricks Agent Bricks rollout looks like
Pick one use case and its data
Choose one agent use case, list the tables and documents it needs, and confirm they are governed in Unity Catalog.
Decide models and residency
Choose models, then decide between pay-per-token (cross-Geo) and provisioned throughput that keeps inference in India.
Build and evaluate
Build the agent in your framework, add AI Search, write judge criteria and run evaluations on real questions.
Trace, budget, harden
Read MLflow traces, set Unity Gateway budgets and guardrails, and tag each endpoint so DBUs map back to teams.
Release and commit
Release behind evaluation gates, forecast DBUs from pilot usage, then settle a committed-use quote in INR with GST.
The review scoreboard
Modelled on Gartner Peer Insights structure. *Counts and breakdowns are illustrative pending verified review collection.
Quick poll — what’s driving your evaluation?
Talk to an advisor“Our claims-triage agent reads policy PDFs through AI Search and every answer is traced, which our model-risk team insisted on.”
“Swapping to a cheaper model was a gateway setting, and the judges confirmed within a day that answer quality had held.”
“Provisioned throughput kept our patient-query bot inside the India Geo, though at low traffic it costs more than per-token.”
“We had planned on Knowledge Assistant, then learnt it no longer takes new customers and rebuilt on Genie Agents. Check first.”
“Serving, search and evaluation each add DBUs, so the first bill was hard to read until we tagged every endpoint by team.”
“GPU serving for our demand-forecast model was quick to stand up, and the data never left the workspace it was trained in.”
Where everyone sits — the grids
Analyst firms bury this view behind paywalls, and review-site grids score satisfaction, not fit for your estate. So here’s TechBag’s synthesis of the enterprise AI agent platform market — tap any vendor to see why it sits where it does.
TechBag AI Agent Platform Grid
Execution strength vs product vision — the classic market map, minus the paywall.
Gartner DSML Leader 2026; from $0.07 a DBU.
Agent Lifecycle × Data Proximity
The grid nobody publishes — how much of the agent lifecycle a platform covers (build, evaluate, serve, govern) vs how close it runs to your governed enterprise data.
Build, evaluate, serve and govern next to lakehouse data.
Positions are TechBag’s illustrative synthesis of public review-platform data and vendor documentation — not a reproduction of any analyst graphic. Verify before relying on it.
Databricks Agent Bricks vs the AI agent platform field
Against Snowflake Cortex AI, Salesforce Agentforce, Amazon Bedrock AgentCore, Google Agent Runtime and Microsoft Foundry — on models, building, evaluation, retrieval, governance, price and India.
| Dimension | Databricks Agent Bricks | Snowflake Cortex AI | Salesforce Agentforce | Amazon Bedrock AgentCore | Google Agent Runtime | Microsoft Foundry |
|---|---|---|---|---|---|---|
| What it is | Databricks’ AI umbrella | AI inside Snowflake | CRM agent platform | Agent runtime on AWS | Managed agent runtime | Azure AI platform |
| Where it runs | Databricks workspace | Snowflake account | Salesforce cloud | Any framework, on AWS | Google Cloud only | Azure only |
| Model choice | Eight model families | Several labs, in credits | Salesforce-managed | Bedrock catalogue | Gemini plus Model Garden | Large model catalogue |
| Agent building | Code-first, open source | Cortex Agents | Low-code Agent Builder | Framework-agnostic | Agent Development Kit | Foundry Agent Service |
| Evaluation and tracing | Judges plus MLflow | AI Observability | Testing in Salesforce | AgentCore Evaluations | Gen AI evaluation | Foundry evaluations |
| Retrieval | AI Search | Cortex Search | Data 360 grounding | Bring a store | Google search services | Azure AI Search |
| Custom models and GPUs | GPU serving, AI Runtime | Snowflake ML | Not its purpose | Via SageMaker | Training on GPUs, TPUs | Fine-tune and deploy |
| Governance | Unity Catalog + Gateway | Horizon + AI Gateway | Trust layer, CRM rules | Identity and gateway | Google Cloud IAM | Azure controls |
| Business context | Lakehouse data | Snowflake data | CRM-native | You wire it | Google data services | Azure and Microsoft data |
| Pricing model | DBUs per product | AI credits per token | Conversations or credits | Per vCPU, GB and call | Per vCPU and GiB-hour | Per service |
| Published entry price | $0.07 per DBU | $2.00 per AI credit | ~$2 a conversation | $0.0895 per vCPU-hour | $0.085 per vCPU-hour | No single figure |
| India data region | Partly in-Geo | Mumbai and Pune accounts | Hyperforce India | Mumbai, Hyderabad listed | Mumbai listed | Varies by model |
| Lock-in and exit | Open MLflow, own endpoints | Snowflake-bound | CRM-bound | Code portable | ADK is open | Azure-bound services |
| Best fit | Lakehouse-first AI teams | Snowflake-centred teams | Front-office agents | AWS builders | Google Cloud shops | Microsoft estates |
Which approach fits you?
Honest fit signals — because the fastest way to lose your trust is to pretend one product wins every scenario.
Choose Databricks Agent Bricks if…
- ✓Your data already lives in Databricks and agents must inherit Unity Catalog permissions and lineage
- ✓You want many model providers, evaluation and tracing on one DBU invoice rather than several AI contracts
- ✓Provisioned throughput in Mumbai or Pune meets your residency needs for production inference
Compare alternatives if…
- ✓Your agents act mainly on CRM records — Salesforce Agentforce is built around them
- ✓Your data sits in Snowflake — Cortex AI keeps agents beside it, paid in AI credits
- ✓You want a framework-neutral runtime on one hyperscaler — AgentCore, Google’s Agent Runtime or Foundry
Do not expect…
- ✓Knowledge Assistant or Supervisor Agent for a new account — both are closed to new customers
- ✓Every model call to stay in India — pay-per-token needs cross-Geo processing
- ✓A flat monthly price — serving, search, evaluation and GPUs each meter DBUs
TechBag has no ai agent platforms guide yet, so Databricks Agent Bricks sits outside the category guides. Browse all products to compare it with the rest of the catalogue. →
What does stitching AI services together cost you?
Drag the sliders (agents in production; ML-engineer hour cost). Estimates model the engineering time spent wiring separate model, retrieval, evaluation and tracing services together and keeping their permissions in step, at an assumed 20 hours per agent a month, with 50% of it removed by one governed platform. Both figures are assumptions. Illustrative.
Loaded cost = salary + overheads per productive hour. Illustrative only — your TechBag quote models your actual environment and modules.
Three ways to consume it
Published per DBU, USD list, billed per second for each underlying product. Serverless real-time inference: $0.07 a DBU in US East, $0.074 in AWS Mumbai and Azure Central India. CPU Model Serving: 4 DBU an hour (about $0.28 in US East). AI Search Standard: $0.28 a unit-hour ($0.30 in Mumbai). AI Runtime A10 GPU: $2.50 an hour ($3.27 AWS Mumbai). In-Geo model endpoints add 10%. TechBag quotes in INR with GST.
Pay-per-token models
Best for pilots and spiky traffic
- $0.07/DBU serverless inference, US East
- Llama 3.3 70B ≈ $0.50 in, $1.50 out per M
- From India, needs cross-Geo processing
+ Platform add-ons
Best for a broader rollout
- Scoped to your estate
- Add-on modules as needed
- Phased, right-sized deployment
Provisioned throughput
Best for steady load that must stay in India
- Runs inside the India Geo
- Reserved capacity, billed in DBUs
- In-Geo endpoints add 10% to the rate
Buy it for less — TechBag pricing beats list
Whatever the list prices above, TechBag negotiates a significantly better deal — with GST-compliant INR invoicing and local support. Ask us for your discounted quote.
Get an India-ready quote
Tell us your requirements and current tools — we’ll model it against what you spend today.
The 8 questions to ask every vendor
Take this into your next vendor call — including ours.
Which agent use case comes first, and which tables and documents must it read through Unity Catalog?
Must inference stay in India? If so, budget provisioned throughput, since pay-per-token needs cross-Geo processing.
Which hosted models will you use, and are any of them external providers billed outside Databricks?
Does anything you were shown rely on Knowledge Assistant or Supervisor Agent, now closed to new customers?
Which judge criteria and test questions will gate a release, and who signs off the evaluation results?
Which pieces are Beta or Public Preview, such as AI Runtime and managed memory, and is that acceptable here?
What DBUs did the pilot use across serving, search and evaluation, and how will Unity Gateway budgets cap it?
What committed-use discount applies, and can TechBag itemise the Agent Bricks quote in INR with GST?
Questions buyers ask
Ready to evaluate Databricks Agent Bricks?
Describe the agent you want and the data it needs, or let a TechBag advisor settle the residency model, run a governed pilot and get the quote itemised by SKU in INR.
Stats, ratings, review counts and pricing are illustrative and sourced from public materials; verify before purchase.