Enterprise AI agents are no longer limited by model intelligence alone; the real challenge is governed production deployment. That’s the message behind a new partnership between Databricks and OpenAI, which shifts the conversation from model benchmarks to platform architecture. Instead of another announcement about raw capability, the focus is on a set of Databricks agent tools designed to solve the practical hurdles of running AI agents at scale. These tools target five specific areas: context management, permissions, lineage tracking, cost controls, and evaluation. For anyone involved in enterprise AI agent deployment, this signals a move toward treating production agent infrastructure as a first-class concern. The idea is simple: model intelligence is no longer the only deployment constraint, and Databricks is selling these five capabilities as the missing production layer around frontier models. If you’ve been wrestling with how to deploy governed AI agents without losing control over data access or spending, this announcement offers a concrete path forward.
1. Databricks Agent Tools: Governing Codex and GPT Agents via MCPs
The flagship offering in this slate directly answers the question of how to grant AI agents data access without sacrificing security. Databricks Agent Tools uses the Model Context Protocol (MCP) to give Codex- and GPT-powered agents governed access to your enterprise data. MCP acts as a standardized interface, but with a critical twist: it routes every request through your existing governance controls. This means every data call an agent makes is subject to the same fine-grained permissions and data lineage tracking you already rely on for human users and ETL pipelines.
This setup directly addresses a key evaluation question: can your existing lakehouse governance, Unity Catalog permissions, and AI gateway policies be reused for agent fleets? The answer here is yes. You do not need to build a separate permission system for AI. By reusing Unity Catalog roles and policies, Databricks Agent Tools ensures that MCP governed agent access behaves exactly like your current data platform. For administrators, this represents a practical approach to Codex agent data governance and GPT enterprise agent security. The enforcement layer is already in place, and now it simply extends to every autonomous query the agent makes.
2. Agent Bricks Platform: Unified Infrastructure for Agent Fleets
Moving beyond individual agent security, the Agent Bricks platform tackles the broader challenge of managing entire agent fleets at scale. It is a unified AI platform designed to reduce fragmentation across your agent stacks, providing a centralized infrastructure for model choice, enterprise context, tool access, evaluation, monitoring, memory, deployment, and governance. With Agent Bricks, you get a single system to handle the full agent lifecycle, from development to production. This is crucial for enterprises building multi-model agent fleets, such as those using GPT and Codex together. The platform supports running these different models side by side, making agent fleet management more consistent and controllable. By centralizing these capabilities, you simplify scaling AI operations while ensuring data governance and compliance remain integrated from the start. Agent Bricks is a practical addition to the Databricks agent tools, offering the infrastructure needed to move from isolated agents to coordinated, enterprise-wide deployments.
3. Unity AI Gateway: Policy Layer for Agent Traffic and Costs
After setting up coordinated agents with Agent Bricks, the next step is controlling how they operate. That is where Unity AI Gateway comes in as the policy layer for your agent traffic. It acts as the control plane, managing permissions, rate limits, and cost tracking for every agent you deploy. With Unity AI Gateway, you can enforce policies for which tools an agent can call and which models it can access. This keeps your agents from overstepping their intended scope. The gateway also allows you to layer these policies on top of existing Unity Catalog permissions, so agent access aligns with your broader data governance rules. This is crucial because production agent quality increasingly depends on context, permissions, evaluation, and cost controls around the model.
Cost controls for production agents are a standout feature. You can set token limits per agent or define budget thresholds that stop traffic when costs exceed a certain point. This prevents runaway spending from unexpected usage patterns. Reusing existing gateway policies streamlines setup, as you can apply the same rules you already use for other AI services to your Databricks agent tools. By integrating agent permission policies and AI gateway cost controls, Unity AI Gateway provides practical agent traffic governance without adding management overhead. You get a single place to monitor and adjust how your agents consume resources, making production deployments more predictable and manageable.
4. Governed MCP Access: Enterprise Data Integration for Agents
After getting agent traffic under control, the next challenge is managing the data those agents can reach. Databricks Agent Tools tackles this by implementing the Model Context Protocol (MCP) for Codex and GPT-powered agents. MCP gives agents a governed way to query enterprise data in real time, directly from Unity Catalog and other connected sources. This turns enterprise agents into a governed data-access problem, which is exactly how the Databricks-OpenAI update frames it. Instead of letting agents roam freely, you set fine-grained permissions on what data each agent can see. Lineage tracking also records every data query, so you maintain full visibility into how agents use your information assets.
MCP’s governed data integration extends beyond the Databricks lakehouse itself. It supports external enterprise data sources, meaning your agents can tap into systems you already rely on without complicated custom connectors. For each integration, you define access rules through Unity Catalog’s existing permission model. The result is a single control point for all agent data queries across your organization. This practical Model Context Protocol implementation helps you scale enterprise agent deployments confidently, knowing that sensitive data stays protected while agents still get the timely information they need.
Also worth a read: Salesforce Data Thefts Continue via Klue App.
5. Evaluation and Monitoring Tools for Production Agent Quality
Once you have secured your agent deployments with proper context and permissions, the next step is ensuring they actually perform well in production. Databricks agent tools now include built-in evaluation and monitoring capabilities that help you track agent reliability and safety without guesswork. As Databricks highlighted in their July 6 DAIS 2026 partnership recap with OpenAI, production agent quality increasingly depends on context, permissions, evaluation, and cost controls around the model. That means you need more than just a working prototype — you need real-time visibility into how your agents behave once they are live.
Built-in evaluation for agent outputs. You can measure agent accuracy, latency, and safety using standardized agent evaluation metrics. These metrics help you catch issues like hallucinations, slow responses, or unsafe outputs before they affect users. Instead of relying on manual checks, you get automated scorecards that flag problems immediately. Monitoring agent fleets in production. Dashboards give you production agent observability — a live view of how each agent is performing across your organization. You can see error rates, response times, and even agent cost analysis to tie usage back to your budget and compliance requirements. This combination of evaluation and monitoring ensures your agents stay reliable, efficient, and safe as you scale.
Frequently Asked Questions
How does Databricks Agent Tools enforce governed access to enterprise data through MCPs?
Databricks agent tools use Model Context Protocols (MCPs) to securely connect agents to enterprise data sources. These MCPs enforce access by inheriting your existing Unity Catalog permissions and AI gateway policies. This means agents can only retrieve data that you, the user, are authorized to see, ensuring governed access without manual reconfiguration.
Can we reuse our existing Unity Catalog permissions and AI gateway policies for agent fleets without changes?
Yes, you can. Databricks agent tools are built to work directly with your existing Unity Catalog permissions and AI gateway policies. You do not need to create separate rules for agent fleets. This reuse saves time and keeps your governance consistent across all AI interactions.
How do cost controls and evaluation work in practice for production agents?
Databricks agent tools include cost controls that let you set spending limits per agent or fleet. Evaluation is handled through integrated monitoring dashboards that track usage, performance, and costs. You can review these metrics to optimize your deployment, giving you a practical way to manage production agents.
