As enterprises adopt agentic AI, they need to shift from reactive systems of intelligence to proactive systems of action to equip the agents they’re building with the context and performance they need, plus regulator-grade accountability, where every decision is explainable and auditable.
At Google Cloud Next ‘26, we discuss how our Agentic Data Cloud enables a system of action, and Yahoo’s digital media buying platform is a compelling example of this vision. Yahoo partnered with Google Cloud to build its Seller Agent digital media buying platform using Google Data Cloud graph technologies. Seller Agent condenses multi-week manual processes into fully governed, live campaigns that can be executed in just seconds. Ultimately, this agentic platform serves as a powerful blueprint for multiple industries, demonstrating that autonomous systems can operate at remarkable speed while remaining strictly accountable.
"Yahoo's mission is to be a trusted guide through the digital world. In partnership with Google Cloud, we're extending that promise to advertisers: agentic media buying that's fast, transparent, effective, and built to be trusted." - Gabriel DeWitt, Head of Monetization, Yahoo
In this blog, we explore the shift toward agentic AI, examine how Yahoo’s Seller Agent architecture solves for speed and trust in media buying, and show you how to apply this graph-based pattern to build trusted systems of action in your own organization.
For years, complex, high-value workflows—like premium digital advertising campaigns—have required weeks of human handoffs, fragmented spreadsheets, and manual analysis. Yahoo recognized that agentic AI could collapse this timeline, allowing agents to plan and execute campaigns in mere seconds. This leap from manual to autonomous execution represents a massive opportunity to reclaim operational efficiency and ensure more of every dollar reaches measurable outcomes. But simply dropping LLMs into a high-stakes workflow does not solve the problem; an agent attempting to negotiate contracts or ad placements without a deterministic understanding of real-time inventory, pricing rules, and business constraints is prone to hallucinate — potentially resulting in disastrous deals. A trusted agentic platform requires a definitive, real-time source of truth, ensuring it acts on hard facts rather than statistical guesses.
Furthermore, speed and factual grounding are only half the equation. The moment an AI agent starts moving real budgets, it faces scrutiny from regulators who demand instant answers to why specific decisions were made or which policies were applied. Digging through raw system logs after the fact is the wrong control surface for autonomous execution. Real-world systems of action require regulator-grade governance and auditability built directly into the workflow, not bolted on as an afterthought.
Yahoo's mission has always been to be a trusted guide through the digital world. Agentic media buying extends that promise to advertisers, agencies, publishers, and regulators who entrust Yahoo with their budgets — and expect real accountability. The issue was automating campaign execution in a way that was explainable, governable, and auditable.
To meet this challenge, Yahoo built its Seller Agent as a multi-agent system running on Google Cloud. Buyer requests enter through a planning supervisor agent running on Google Kubernetes Engine (GKE) and orchestrated with Google's Agent Development Kit (ADK). The supervisor decomposes each request into specialized tasks including inventory discovery, audience matching, forecasting, pricing analysis, package recommendation, governance review, and execution. Agents coordinate through the open Agent2Agent (A2A) protocol, while Gemini Enterprise Agent Platform hosts models for embeddings, forecasting, and graph learnings.
But the true breakthrough — what makes autonomous execution both fast and fully transparent — is the platform’s dual-graph foundation. The platform is anchored by two specialized graph systems with an intentional separation of duties: a knowledge graph that’s optimized for acting, and a second context graph for remembering and learning.
Powered by Spanner Graph, Yahoo’s knowledge graph represents its monetization business as a connected operational model, grounding every agent decision in business reality. It models advertising products, placements, audience segments, inventory, contracts, and governance controls as first-class entities and relationships. Crucially, policies live directly within the graph as versioned relationships rather than being buried in application logic. This design allows the system to evaluate products, contractual obligations, consent requirements, and regulatory constraints together in a single, unified graph traversal.
The graph acts as a semantic contract across the agentic platform. During campaign evaluation, an agent can navigate from initial buyer requirements to eligible audiences and governing policies within a single query plan. Gemini Enterprise Agent Platform embeddings enrich these entities with semantic similarity, while graph neural networks contribute inferred relationships. Ultimately, this allows agents to do more than just retrieve available inventory — they understand exactly why it is relevant and help ensure it satisfies all governing constraints.
Execution at agent-scale is only safe if it is entirely transparent — which is the core function of the context graph. Every time the Seller Agent takes an action, that exact operational span is captured by the BigQuery Agent Analytics plugin. In addition to logging the raw events, the system shapes this evidence into a typed, queryable context graph using BigQuery Agent Analytics SDK utilizing Yahoo's decision-trace ontology, stored in BigQuery Graph.
Consequently, every decision point, candidate package, policy evaluation, specialist-agent delegation, and execution outcome becomes a connected graph of evidence. Because this trace is structured as a typed graph, explaining the agent’s decision making process becomes a simple query. An auditor can instantly trace a decision from the originating campaign brief through every score that’s assigned and policy that’s applied. This transforms autonomous behavior from an opaque process into a fully transparent and continuously improving record of decision-making, helping to ensure absolute accountability.
For a concrete example of the architecture in action, consider an ad campaign run. What traditionally required weeks of coordination across planning, sales, operations, and compliance can now be completed in seconds through two simultaneous processes. **Acting via the knowledge graph. **This pipeline moves the budget, navigating linearly from the buyer's request to a live campaign ground on the knowledge graph. This proceeds in four steps:
Submitting the brief: A buyer agent submits a campaign brief over Ad Context Protocol (AdCP) that describes the desired audience, budget, geography, and business objective.
Knowledge retrieval: The Seller Agent queries the knowledge graph to identify relevant inventory, audiences, contractual availability, historical performance, and governing policies.
Evaluation and scoring: The agent evaluates these factors together to assemble a package of media buying candidates. Forecasting models score the opportunities, while a governance agent independently reviews consent, brand safety, and regulatory constraints.
Approval and execution: The package is either approved automatically under policy thresholds or escalated for human review. Once approved, the media buy is executed and activated.
Auditing and learning via the context graph. While the execution pipeline moves forward, this parallel loop continuously captures the system's reasoning in the context graph, helping to ensure transparency and improve future cycles. This offers the following capabilities:
Continuous capture: Every candidate considered, score assigned, policy applied, and governance decision becomes a connected record in the context graph, linked to the originating campaign session.
Closed-loop learning: As delivery, attribution, and outcome signals arrive, they are joined back to the decisions that produced them, creating the training data that improves future recommendations.
Instant explainability: If an advertiser asks why a particular package was selected or which policies influenced the outcome, the answer is preserved in the context graph and reachable through a single query.
The result is a platform where knowledge, decision-making, governance, measurement, and learning operate together — allowing autonomous media buying to remain explainable, auditable, and continuously improving.
The era of AI as a mere advisor is ending. Enterprises are demanding systems of action — autonomous agents capable of executing complex, multi-step workflows. But in regulated sectors, the speed that AI brings to the table turns into a liability if you cannot prove how a decision was made. The primary barrier to autonomous execution is no longer intelligence; it is trust.
The architecture that Yahoo and Google Cloud built provides a broadly applicable blueprint with which to solve this. While designed to fix the bottlenecks of digital media buying, the underlying pattern applies to any industry managing high-stakes decisions — from financial trading to supply chain logistics. To operate at agent speed but still maintain human oversight, enterprises must adopt a new architectural baseline that:
Grounds decisions in business reality: Agents cannot rely on probabilistic models alone. They must be grounded by a knowledge graph that deterministically maps your business logic, active contracts, and compliance rules.
Builds an auditable memory: You cannot govern what you cannot trace. Every agentic action must be captured in a context graph, creating an immutable, queryable record of exactly why a decision was made and which alternatives were rejected.
Embraces open interoperability: Trust requires transparency. By building on open protocols and provenance standards, industries can establish a common, auditable language for agentic behavior.
As foundational models become commoditized, enterprises’ competitive advantages are shifting. Long term, your moat will not be the language model you deploy, but the proprietary graph of your business operations and governed history. Likewise, the future of enterprise AI isn’t simply systems that can act, but systems that can explain, govern, and take accountability for those actions.
Ready to build your own trusted system of action? Start by exploring Spanner Graph to ground your agentic workflows in business reality. Next, use BigQuery Graph to build an auditable memory that powers closed-loop learning and regulator-grade explainability. You can begin capturing and analyzing these operational traces today using the BigQuery Agent Analytics Plugin and SDK. Finally, review the Ad Context Protocol to understand the open communication standards underpinning Yahoo’s agentic platform.