01The 99% thesis. Why the plumbing, not the model, decides which agents reach production.
02This week, in brief. Genie's free tier and budgets, Feature Views in preview, SQL alerts' compliance default, and the DABs release wave.
03From Brickster.ai: Your curated Databricks digest for the week.
01
🧱 The 99% Thesis
AGENT PLATFORM
Your demo took a weekend. Production takes a year.
Agent Bricks is built on one uncomfortable number: the model is 1% of the work.
By now, every team has built an impressive agent demo. The core loop — an LLM that reasons, calls a tool, and responds — is a weekend project. Databricks' framing is blunt: that loop is roughly 1% of the work. The other 99% is the unglamorous plumbing that decides whether the thing ever survives contact with production.
Token capacity. Deployment. Security. Evaluation. Monitoring. Memory and context. Sharing and governance. None of it shows up in the demo, and all of it is where enterprise AI projects quietly die.
Agent Bricks is Databricks' bet that whoever owns the 99% owns enterprise AI. And at last month's Data + AI Summit the platform posted the kind of numbers that make the bet look less like a slogan: 100,000+ agents built and 1+ quadrillion tokens processed per year.
⚖️ The split nobody budgets for
The 1%
The core loop
Prompt the model
Call a tool
Stream a response
The part everyone demos. A weekend of work.
The 99%
Everything that ships it
Token capacity & cost control
Security, sandboxing & governance
Evaluation & monitoring
Memory, context & state
Deployment, versioning & sharing
The part that decides if you reach prod. A year of work — or a platform.
⚙️ How Agent Bricks covers the 99%
Composition
🔀 Any model, any harness
Model choice now spans OpenAI, Anthropic, Gemini, Qwen, Kimi, and Grok (via SpaceX — the former xAI). Bring your own framework too: LangGraph, CrewAI, Agno, and the Claude Code and OpenAI Agent SDKs are all supported. No rewrite to switch.
Control
🔒 Governance & secure execution
Agents, tools, and models register in Unity Catalog and route through the Unity AI Gateway for access control, cost limits, and audit. Databricks Sandbox (Beta) runs agent code in isolated, secure environments so an agent's actions stay contained.
Context
🧠 Managed memory on Lakebase
Agents get managed memory (Beta), powered by Lakebase (managed Postgres) under the hood — long-term memory out of the box, with full short-term session state if you run your own Lakebase instance. Context persists across sessions, so agents stop starting from zero every conversation.
Quality
📏 Evaluation with the CLEARS framework
Agent quality gets a standardized scorecard in MLflow — CLEARS: Correctness, Latency, Execution, Adherence, Relevance, and Safety. You cannot ship what you cannot measure.
🆕 This week: the 99% grew a memory
Omnigent's contextual policies: governance that remembers the session
The clearest proof of the thesis this week isn't a model release. On July 6, Databricks shipped contextual policies in Omnigent, its open-source (alpha) meta-harness that wraps coding agents like Claude Code and Codex. Policies can now read session state and gate the agent's next action on what it has already done, instead of applying a static allow/deny list.
The built-in examples read like a security team's wish list: a Bell-LaPadula no write-down rule that stops an agent from copying confidential content into less-restricted files, a running risk score that flips sensitive actions to ask-a-human past a threshold, per-session spend caps, and intent-based least privilege that locks a session to the tools relevant to its original prompt.
This is the 99% in its purest form. Nothing about the model changed on July 6. What changed is whether you can let an agent act without a human watching every step — the actual blocker in most enterprises.
100k+
Agents Built
1+ Quadrillion
Tokens / Year
1%
Is the model
In production, not slideware
Agents built on the platform are shipping at AstraZeneca, 7-Eleven, Fox Corporation, and Block — pharma, retail, media, and fintech. The spread matters as much as the logos: same 99% problems everywhere.
🔭 What this means for your team
1. Budget for the 99%, not the demo. If your agent roadmap is mostly prompt-and-model work, you've budgeted for 1% of the project. Evaluation, governance, and memory are the real line items.
2. Don't over-index on the model. Any-model support means model choice is now a config detail, not an architecture decision. Optimize the platform around it, not the other way around.
3. Static allowlists won't survive agents. Omnigent's contextual policies are an early look at governance that tracks session state. If your agents touch confidential data, the no-write-down example is worth ten minutes.
4. Watch the lock-in trade. Owning the 99% is powerful for Databricks and convenient for you. Just enter it knowing the platform, not the model, is the thing you're committing to.
Four more items worth your attention from the past week:
PRICING
Genie now bills on LLM DBUs — with a free tier
As of July 8, Genie Agents (formerly Genie Spaces), Genie Code, and Genie One share one pay-as-you-go model billed on underlying LLM usage in DBUs, with a 150-DBU per-user monthly free allowance (~$10.50 at US East rates). Compute that runs the queries, such as a SQL warehouse, is billed separately. Account admins can set budgets and alerts at the account, workspace, user-group, or individual level, so the free tier isn't a silent overage trap.
ML
Feature Views: define a feature once, use it everywhere
Now in Public Preview, Feature Views is a managed feature-pipeline framework where one definition powers both training data and batch or real-time inference, eliminating training/serving skew by construction. Materialized features live as governed Unity Catalog objects. Databricks claims Kafka-sourced streaming features served at 200ms end-to-end p99, but the streaming path needs an Enterprise-tier workspace in a Lakebase-supporting region.
COMPLIANCE
SQL alerts are now default-on in compliance workspaces
As of July 9, Databricks SQL alerts are available by default in workspaces with the compliance security profile enabled, across all the CSP standards Databricks supports — HIPAA and FedRAMP included. A small unlock on its own, but it's exactly the kind of default that decides what regulated teams can actually run.
CI/CD
The DABs-anchored tooling wave keeps rolling
CLI v1.7.0 (July 9) is mostly a Databricks Asset Bundles release: profile selection now takes precedence over auth env vars, plus PyDABs permission fixes and better handling of Unity Catalog catalog/schema drift. It lands alongside the same-week Delta Lake 4.3.1 patch and SDK Go/Java refreshes. Individually minor; together they're the plumbing making DABs the default declarative CI/CD path for Databricks deploys.
03
🧱 From Brickster.ai
This is the short version. The full week-of-July-13 digest goes deeper on each of these — Feature Views' streaming path, the Genie budget mechanics, and the rest of the release wave — over at brickster.ai/digest.
Quick links this week:
→ brickster.ai/news — This week: Genie's new billing model, the Feature Views preview, and Omnigent's session-state governance, in context.
→ brickster.ai/releases — CLI v1.7.0, Delta Lake 4.3.1, and the SDK Go/Java refreshes — version numbers and what actually changed.
→ brickster.ai/assistant — Ask what the 150-DBU Genie allowance means for your workspace, or how Feature Views handles streaming.
🤖 Building an agent this quarter?
The Brickster Assistant searches our full archive and answers with citations.
