Databricks Aims for Agentic Media Buying Scale

Databricks is tackling the complex coordination inherent in media buying with a new architectural blueprint for agentic systems. The company argues that scaling autonomous buyer and seller agents requires more than just advanced AI; it demands a robust foundation for managing data, trust, and operational visibility. StartupHub.ai data notes Databricks holds a strong StartupHub score of 82/100, positioning it favorably against competitors like Snowflake (72/100) and Palantir (85/100).

The core challenge in current media buying, according to Databricks, isn't a lack of talent but the sheer inefficiency of manual coordination. Billions of dollars move daily through fragmented processes involving emails, spreadsheets, and phone calls. This leads to delays and decisions based on outdated information, missing optimal opportunities.

Agentic workflows promise to automate this manual coordination. This shift allows advertising professionals to focus on strategy, targeting, and creative elements. The rise of AI agents capable of pursuing goals, making decisions, and transacting on behalf of users is key to this transformation.

The Need for Standards and Infrastructure

While AI and protocols like MCP can automate the transaction loop, industry-wide adoption hinges on standards. IAB Tech Lab's Agentic Advertising Management Protocols (AAMP) provide a shared vocabulary for inventory and audiences, a common transaction protocol, and a registry model for discovery and trust. These standards enable any compliant buyer agent to transact with any compliant seller agent, akin to how any web browser can access any website.

However, the critical infrastructure where these agents operate, their state, models, identity, and governance, is where Databricks sees its opportunity. The company has developed a reference implementation for agentic media buying on its platform, enabling autonomous buyer and seller agents to discover each other, agree on terms, and close deals.

This solution is built on the official open-source IAB Tech Lab Software Development Kit (SDK), ensuring no vendor lock-in at the protocol layer. It's available as a single-command accelerator, simplifying deployment for organizations.

An Ecosystem for Autonomous Transactions

An agentic media buy involves three primary players: a buyer with a campaign brief, budget, and price ceiling; sellers with ad inventory and pricing catalogs; and a registry for discovery and trust. The transaction itself is a rapid loop of discovery, pricing, and deal closure.

For these agents to operate effectively, a comprehensive system is needed. This system must manage:

• State: Persistent and transactionally correct storage for briefs, identities, catalogs, quotes, and orders.
• Models Served: Governed foundation models accessible by agents for specific tasks.
• Governed Data: Audience and catalog data that agents reason over, with strict access controls.
• Identity and Trust: Verification of agent identity and permissions.
• Observability: End-to-end tracing of decisions and tool calls for debugging and auditing.

Databricks posits that its platform provides this integrated system, eliminating the need to assemble disparate solutions from multiple vendors.

Architecting the Agentic Workflow

The Databricks solution architecture separates the communication protocols between parties from the operational platform where agents run. Each party operates a self-contained application built on Databricks Apps, managing its own state, identity, and models, interacting only through defined protocols.

The buyer agent, for instance, is a hierarchical "crew" of specialist agents using CrewAI. This setup includes a Portfolio Manager for strategy and budget allocation, Channel Specialists for specific media types, and tactical workers for audience planning and execution. These agents leverage Databricks Foundation Model APIs and can be configured to use various LLMs, like Claude models, via the Databricks Unity AI Gateway, allowing easy LLM switching.

Foundational Components on Databricks

For managing transactional state, Databricks Pushes Agentic AI with Lakebase, its serverless Postgres offering, runs alongside the lakehouse. This provides fast, transactional reads and writes for agent operations, autoscaling to handle fluctuating demand.

Governance is layered through Unity Catalog. Data flows from Unity Catalog tables into Lakebase for agent use and is synced back as transactional state. This ensures a governed data round trip with lineage tracking via Databricks managed pipelines, keeping the lakehouse as the source of truth.

Identity and trust are managed natively using OAuth and Databricks' service principals, replacing the need for external API key management. Trust tiers assigned via a registry dictate what data and transaction capabilities a buyer has access to.

Observability is integrated via MLflow tracing, capturing every agent's reasoning steps, tool calls, and decisions. This is crucial for productionizing, debugging, and auditing autonomous systems.

This integrated approach aims to provide the necessary operational foundation, enabling agentic media buying to move from experimental projects to scalable, production-ready deployments.