Data-Native AI Agents: Why Agents Must Move to Your Data

2026-07-15 · Source: Databricks · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure, Cybersecurity & Data Privacy · Depth: Advanced, long

Summary

This post advocates for a "data-native" approach to AI agent architecture, contrasting it with the common "external agent" paradigm. External agents, which pull data into separate AI stacks, introduce significant challenges including governance weaknesses, increased latency from multi-hop operations, fragmented costs (egress, duplicate storage, per-token pricing), complex vendor coordination, and scattered observability. A critical flaw of external agents is their reliance on post-hoc governance, which fails when agents compute on data, leading to incomplete policy enforcement and higher token consumption. Data-native agents, conversely, operate within the existing data platform, unifying governance, security, and observability. Databricks implements this with Unity AI Gateway for traffic control, Unity Catalog for asset management, MLflow 3 for tracing, Model Serving for inference, AI Search for retrieval, and Lakebase for governed agent state and memory, ensuring policy enforcement at query planning time.

Key takeaway

For AI Architects or MLOps Engineers scaling enterprise AI agents, prioritize integrating them directly into your existing data platform. This approach ensures unified governance, security, and cost control, preventing the fragmentation and compliance risks associated with external AI stacks. You should inventory current external AI components to identify immediate governance gaps and begin treating AI primitives as governed assets within your data ecosystem.

Key insights

AI agents must operate within existing data platforms for unified governance, security, and cost efficiency, preventing issues from external, fragmented stacks.

Principles

• Data has gravity; move compute to data, not data to compute.
• Governance must be enforced at query planning, not post-hoc.
• Agent state and memory require the same governance as data.

Method

Integrate AI agents as native workloads within your data platform, leveraging existing governance, security, and observability controls for data, models, tools, and agent memory.

In practice

• Inventory external AI components running outside your perimeter.
• Treat AI primitives as Unity Catalog assets.
• Route all agent traffic through a central gateway.

Topics

• AI Agents
• Data Governance
• Lakehouse Architecture
• Unity Catalog
• AI Gateway
• Agent Memory

Best for: CTO, VP of Engineering/Data, AI Product Manager, AI Architect, MLOps Engineer, Director of AI/ML

Related on AIssential

• The Agentic Enterprise Just Got Real — AI's primary challenge is context management, not intelligence, requiring robust governance for agentic systems.
• Many Companies Use AI. Few Know How to Build an AI-Native Enterprise Data Platform. — AI agents transform enterprise data ecosystems by autonomously executing multi-step tasks, moving beyond simple chatbot interactions.
• Endorsement from a NYSE-listed Company: Why Enterprise AI Must Return Data Sovereignty to the… — Enterprise AI must adopt a sovereignty-first architecture, controlling data movement at the device layer to manage agent behavior and ensure auditability.
• Every Agent Needs a Box — Aaron Levie, Box — Secure, sandboxed environments are crucial for AI agents to operate effectively and safely within enterprise data systems.
• Governing AI agents at scale with Unity Catalog — Effective AI agent governance requires controlling access and monitoring actions, not just reviewing potential behaviors.
• #368 AI Agents Are Now Your Database's Main User | Reynold Xin, Co-Founder at Databricks — AI agents are fundamentally reshaping database architecture, driving the convergence of transactional and analytical systems for real-time, governed data access.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Databricks.