Pyrana vs Databricks Agent Bricks

Architecture

Agent Bricks and Pyrana start from different centers. Agent Bricks puts agents next to the data: agents run on Databricks Apps, read Unity Catalog tables and indexes, and are registered and traced in the same catalog. Pyrana puts the business application at the center: pages, operations, approvals, and versioned outputs, with agents as one mechanism inside, and Databricks as one of several tool sources the application reads from.

Databricks describes Agent Bricks as "the Databricks agent developer platform." Its documentation lists the parts. The Agent Runtime hosts agents built with any framework or harness, stateful or stateless, as their own Databricks app. DurableAgentServer, in the AgentKit library, adds durable execution so runs survive restarts. Databricks Sandbox gives agents an isolated place to run the code they write. Managed agent memory is backed by Lakebase, Databricks' fully managed Postgres. Tools arrive through MCP servers, the Genie One MCP server, AI Search, and AI Functions. MLflow traces each step, and traces can be stored and governed in Unity Catalog.

The scope has widened over 2026. In March, Databricks made the Agent Bricks page the umbrella for all of its AI capabilities. At Data + AI Summit in June it repositioned Agent Bricks as a platform for building agents "with any model and any harness," naming LangGraph, Agno, CrewAI, the Claude Code SDK, and the OpenAI Agents SDK. Databricks' agent framework documentation now sits under Agent Bricks, and the earlier Knowledge Assistant and Supervisor Agent offerings are listed as legacy.

Pyrana's architecture is a platform built once and shared by every application: the Harness (agent runtime), durable workflows on Temporal, tool sources with per-call authorization, tool approval policies, versioned storage, cortIQ for company knowledge, and the App shell with declared pages and a copilot. Identity, authorization, and audit sit under all of it. A client authors one App configuration plus the frontend experience.

Is Agent Bricks a platform for business applications?

It is a platform for building and hosting agents, and Databricks Apps can host the user interface around them. The application logic beyond the agent (who may approve what, how outputs are versioned, how one operation behaves the same from a page, a chat, and a schedule) is code the team writes on Databricks Apps.

Context and knowledge

Agent Bricks grounds agents in lakehouse data and documents: AI Search over unstructured content, Genie Agents configured with trusted data, metrics, and business rules, and Genie Ontology, which Databricks says learns an ontology from data and human-annotated semantics. Pyrana's cortIQ holds company policy and procedure as typed, classified Context Units with evidence, and accounts for what each run retrieved, injected, and cited.

The two address different kinds of knowledge. Genie Agents answer questions about the numbers: what revenue did, which accounts moved. cortIQ holds the rules that interpret them: the capitalization threshold, the approval limits, the close checklist. Each Context Unit carries a claim, its evidence passage with source and location, a knowledge type, a classification that governs how it may change, and a content-hash identity. A regulated unit waits for a person before it changes. Changes supersede rather than overwrite. Citations are content-addressed, so the business context used in any run is provable.

Managed agent memory in Agent Bricks persists session history and context across sessions. Pyrana keeps run outputs as versioned, attributed resources that the next run and the person on the page both read at an exact version.

Agent capabilities

Both run agents with tool calling, multi-step plans, and durable execution. Agent Bricks lets teams bring their framework of choice and adds a sandbox, managed memory, and model access through Unity AI Gateway. Pyrana runs agents on its own Harness, with workflows that fan out to specialized agents per entity and group and publish results back to the application.

In Pyrana a model is assigned per agent through provider adapters (Azure OpenAI, OpenAI, Anthropic, Google Gemini, and others), with the client's own keys. Agents do not write free-form queries: tools are specific to the job, the agent passes the entity and the period, and the tool holds the query. A code interpreter runs deterministic scripts for data shaping, and a data quality check runs before an agent run starts.

Governed actions and approvals

Agent Bricks governs access: which models, MCP servers, and tools an agent may use, with guardrails, rate limits, budgets, and contextual policies written in SQL. Pyrana governs each call: a policy over the call's typed arguments decides whether it runs now or freezes and waits for a named group, and the frozen call is revalidated before it runs.

In Pyrana, when an invocation trips the policy (an amount above a threshold, a missing invoice reference), the platform freezes the operation, exact arguments, scope, and evidence under a hash and queues it to a named approval lane. Reviewer eligibility is checked when the reviewer acts. Dispatch is idempotent, so a retry reuses the receipt.

Does Databricks Agent Bricks have human in the loop?

Databricks documents human feedback for quality: experts label outputs, and MLflow evaluation uses that feedback with LLM judges and custom scorers. We did not find a documented pre-execution approval gate for agent tool calls. Teams that need one build it in agent code and build the routing and records around it.

Building the application

On Databricks, the team writes the agent in code, deploys it with the Agent Bricks CLI, and builds the user interface as a Databricks app in Streamlit, Dash, Gradio, React, or another supported framework; the CLI scaffolds a chat UI. On Pyrana, the team writes one App configuration and designs the frontend experience; pages, the copilot, and authorization already exist.

Genie Agents are the low-code path on Databricks for question answering over governed data. In Pyrana, people can also ask the copilot for a personal page, which it authors, the backend compiles without network or storage access, and the browser renders with every read still authorized.

Governance and security

Databricks extends Unity Catalog from data to AI: agents, tools, and models are registered there, Unity AI Gateway governs model and MCP access, and agent traces flow into the lakehouse. Pyrana applies one eight-step gate to every call from every door, writes the execution row and audit record before dispatch, and fails closed.

Both can run agents as the user. Databricks Apps offers app authorization, where the app's service principal is the identity, and user authorization (Public Preview, enabled by a workspace admin), where the user's downscoped token is forwarded and Unity Catalog applies that user's row filters and column masks. Pyrana keeps the actor chain explicit (an agent with its own identity, an agent on behalf of a user, a workflow as a service principal) and authorizes all three on the same path: tenant, App membership, access profile, scope entitlement. Unauthorized resources read as absent, not denied.

Operations

Agent Bricks runs on Databricks-managed serverless infrastructure, with MLflow tracing and production monitoring. Pyrana runs as SaaS on a dedicated private tenant, or in the client's own infrastructure for large enterprises, with traces, tokens, and tool errors per run available as JSON.

Pyrana releases are explicit: push compiles the App configuration into an immutable version, activate moves the active pointer and is audited, and rollback is an activation of a prior version.

Cost structure

Databricks is consumption-priced in DBUs, billed per second, with discounts for committed use; Databricks Apps are billed per hour of compute while running, and Agent Bricks has its own pricing page. Pyrana is enterprise platform pricing. With clean data available, a Pyrana application typically takes about a week to stand up, and a first engagement runs about six weeks.

Genie Agents moved to pay-as-you-go pricing in July 2026, and Databricks has made Genie One and Genie Agents usage free through January 31, 2027. Model tokens, Apps compute, and storage are billed as they are used.

What it takes to ship the same application

Reference application: an approval-gated procurement request. A person or agent drafts a purchase request; amounts above a threshold need a finance approver; the request waits durably for the decision; everything is recorded.

Using them together

The most common pattern is Databricks as the data layer and Pyrana as the application layer. Databricks is a built-in Pyrana tool source. The user's own principal passes through, so Databricks row-level security applies to the person who started the request, and Pyrana's gate admits and audits the call on its side. Two controls hold at once: the warehouse decides which rows the person may see, and Pyrana decides whether the operation may run and who must approve it.

Databricks also cleans the data. Pyrana does not need to store source data; an upstream ETL layer, for example serverless Databricks, can expose a catalog of clean views that Pyrana's job-specific tools query. A pre-run data quality check confirms the inputs before agents start.

Agents meet in both directions over MCP. An agent built on Agent Bricks can call a Pyrana App through the App MCP server, as the user, through the same gate and approval policy as the App's own pages. A Pyrana App can call a Databricks-hosted agent or a Genie Agent exposed as an MCP server or API as a tool, with that call authorized and recorded.

When not to choose Pyrana

• The work is analytics: business users asking questions of governed lakehouse data. Genie Agents are built for that.
• The team is a data science team that wants to write agents in its own framework and keep everything inside one Databricks account and one Unity Catalog.
• You are building an agent into a data product, and the surrounding users, permissions, and approvals already exist elsewhere.
• You need model training, fine-tuning, or reinforcement learning on your own data. Pyrana assigns models; it does not train them.