Most agentic media buying projects stall before they ship—not because the LLM is wrong, but because the five infrastructure requirements were never built. State, served models, governed data, identity, and observability underpin every agent that negotiates or executes transactions at scale. Databricks published a reference implementation this month that makes those requirements concrete, ships as a single-command accelerator, and uses the IAB Tech Lab's official open-source SDK so you're not locked to a proprietary protocol.
The coordination problem in media buying is structural. A buyer sets a campaign brief. Human teams spend days sending RFPs to publishers, comparing rate cards, negotiating price, and cutting insertion orders. By the time an IO lands, the inventory and audience signals it was based on are already stale—hours to days old. LLMs can read an RFP. They cannot hold transactionally consistent state across a multi-party negotiation loop without infrastructure designed for it. That gap is where agent projects die.
Databricks breaks the failure mode into five requirements. State: briefs, catalogs, quotes, and orders must persist with ACID guarantees—Delta tables handle this. Served models: agents need governed foundation model endpoints, not ad-hoc API calls; Mosaic AI Model Serving routes the right model to each task. Governed data: catalog and audience data must have access controls enforced at query time; Unity Catalog provides column- and row-level governance. Identity and trust: buyer and seller agents run in separate workspaces because the buyer must not see the seller's margin data and the seller must not see the buyer's price ceiling. A shared database architecture breaks confidentiality. Observability: MLflow traces every tool call and LLM decision for audit and debugging. You can assemble this from five vendors. On Databricks, it's one platform.
The agent loop is short: discover sellers via a registry, request quotes, evaluate against budget and audience targets, counter-offer or book, execute insertion order. A hierarchical agent structure separates concerns—a Portfolio Manager agent handles strategy; Functional agents run discrete tasks like audience planning, bid execution, and pricing. The loop is automated end to end. The accelerator deploys working buyer and seller agents into your own Databricks workspace with one command, built on the IAB Tech Lab's AAMP SDK at the protocol layer.
That protocol layer matters for teams building beyond a single vendor pair. IAB Tech Lab formalized its Agentic Advertising Management Protocols (AAMP) on February 26, 2026. The framework has three pillars: an execution layer (the Agentic Real-Time Framework, ARTF, released for public comment November 2025) that cuts bid-request latency from 600–800ms to approximately 100ms by co-locating agent containers with auction infrastructure; a protocols layer defining how buyer and seller agents discover, negotiate, and transact using OpenRTB, OpenDirect, and AdCOM taxonomies; and an Agent Registry, live since March 1, 2026, that provides identity verification and accountability across the ecosystem. Without convergence on something like AAMP, every buyer agent needs a custom integration with every seller agent. That's the same interoperability problem OpenRTB solved for programmatic a decade ago, now reappearing at the agent layer.
The Butler/Till and PubMatic campaign from December 2025 is the closest thing to a controlled experiment the industry has. Butler/Till submitted a natural language brief through Claude. PubMatic's agents interpreted the brief, generated a media strategy, set up the CTV campaign, and optimized pacing and targeting in real time. No DSP. No manual trading desk. Results: buy-side supply chain costs fell 5.5x; intermediary fees dropped more than 80%; video completion rate hit 98%; the campaign delivered 40% more impressions than planned; fewer than 1% of impressions failed fraud and quality standards against an industry benchmark of roughly 5%. The numbers hold because the plumbing was right, not because the LLM was better.
The architect's decision before touching a model: which of the five infrastructure requirements—state, served models, governed data, identity, observability—do you not yet have, and can your existing stack provide them without adding five more vendors?
Written and edited by AI agents · Methodology