Agentic AI initiatives have moved from prototype to production faster than the platforms underneath them. Models can reason, infrastructure can scale, and data exists everywhere. What breaks at production scale is context, and with it trust, confidence, and value. Google’s Agentic Data Cloud, launched at Cloud Next 2026, addresses that gap with a unified architecture built to ground agents in trusted enterprise context. The release spans a universal context engine, an agentic-first practitioner surface, and a cross-cloud lakehouse, with a clear bet on the semantic layer as the new center of gravity in the agentic AI stack.
The first wave of AI agents in the market produced more demos than actual running systems. Across one industry after another, the same arc plays out as pilots work within controlled scope, agents complete the tasks they were built for, yet projects then stall when leadership pushes for the next ten use cases. The technology that worked in the demo does not generalize, and the reason is rarely the model. Every additional use case forces a fresh round of context engineering, governance review, and data plumbing nobody scoped into the original budget.
Agentic AI projects face a long list of challenges, from cultural readiness and skills gaps to runaway costs and legacy stack integration. Three of those challenges sit squarely in the data platform layer, and they show up consistently across nearly every program trying to scale.
Each of these adds latency, cost, and risk to every additional agent. At single-digit pilot counts the friction is manageable. Beyond that, the program grinds to a halt.
Enterprises need business semantics to live close to the data, so every agent inherits the same definitions of things like revenue, customer sentiment, churn, and pipeline that drive the rest of the organization. Agents have to be able to reason over historical data and act on live data in the same step, without copying between systems. The semantic layer itself has to be revisited, evolving from a static definition store into an active inference layer that maps relationships across structured and unstructured assets and reaches the bulk of enterprise data living in documents, transcripts, images, and e-mail. Cross-cloud reach has to come without an egress penalty, because most enterprises now run data across multiple cloud providers by default. And every action an agent takes has to leave the same audit trail a financial transaction would, because no enterprise should roll agents into production without knowing exactly who did what and when.
At Google Cloud Next 2026, the company advanced all six layers of the enterprise AI stack, from custom silicon and frontier models through data, security, tooling, and the agent surface. The Agentic Data Cloud sits at the data management and context layer, and it is the most consequential thread for any enterprise putting agents into production. It organizes around three innovation areas, all focused on providing agents with cleaned, organized data as context for reasoning and analysis.
Beyond the three pillars, Google paired the release with deeper investments in the same architectural threads, with BigQuery Managed Iceberg Tables (GA) anchoring the cross-cloud open lakehouse, BigQuery Graph supporting multi-hop reasoning across business relationships, and a new Spanner Columnar Engine running analytical queries alongside transactional workloads. Performance work followed underneath, with Lightning Engine for Apache Spark, Managed Lustre, a Bigtable in-memory tier, and BigQuery fluid scaling all aimed at the throughput that agent-scale workloads demand.
All of this is only the data-centric slice of Cloud Next 2026, which also brought 8th-generation TPUs (TPU 8t for training, TPU 8i for inference); the Gemini Enterprise Agent Platform that now folds in Vertex AI as the home for building, orchestrating, and governing agents; Gemini 3.1’s advances in reasoning and multimodality; and a new agent security stack riding alongside the Agent Platform.
The packaging inside the Agentic Data Cloud carries more weight than any individual product in the release. Google Cloud has spent more than a decade building services that often read as an à la carte menu, with several product name changes along the way. The same pattern shows up across the rest of the hyperscalers, where sprawling portfolios leave customers asking where to start. Now, with Dataplex and Vertex giving way to new umbrella names, the new vocabulary finally matches what the portfolio is being asked to do in an agent-first world. The Agentic Data Cloud frames the portfolio as one architecture for grounding enterprise agents in trusted, governed business context, giving buyers a coherent answer to where business context lives in an agentic world.
The bet underneath this narrative is that data gravity has moved up the stack. For most of the last decade, the lakehouse won the gravity argument by bringing analytical work to where the data sat. The Agentic Data Cloud assumes agents will not behave the same way. Agents need to understand the business before they can act on its data, and Google has positioned the Knowledge Catalog and semantic layer as the new center of gravity. The “Knowledge Catalog” name is one place where the marketing undersells the product, and it risks reading as a plain catalog when the offering is far more powerful and intelligent, with semantic inference, relationship mapping, and governance well beyond what the term implies. Agents inherit context rather than reconstruct it for every task. Most major platforms are arriving at similar conclusions, but Google’s framing is the sharpest the field has seen to date, and early customers are already putting it to work. Yahoo’s intelligent seller agent, built on Spanner, Vertex, BigQuery, and GKE, is one example of what that framing looks like running in production.
During Google Cloud Next, Mike Leone sat down with Mikul Bhatt of Yahoo to discuss the evolution of digital advertising and how Yahoo is automating the sell side of the business using an intelligent seller agent running on Google Cloud.
Google Cloud has been pushing openness for years, but the expansion of its cross-cloud reach finally makes that story land. Earlier iterations of Google’s stance on openness and open formats arrived unevenly on the data side. The announcements at Google Cloud Next close a meaningful gap. Cross-Cloud Interconnect into the data plane, the Iceberg REST catalog as the connective tissue, and bidirectional federation across Unity, Polaris, and Glue together really change the conversation. The competitive implication is that Google is choosing to compete as the best engine on open formats wherever the data sits, rather than asking customers to consolidate.
Cross-cloud performance under production load is one area Google will need to address. Federation that performs well in benchmarks but slows under petabyte-scale agent workloads risks pushing customers back to copying data, and the unified story relies on that not happening. Catalog inference quality on unstructured data is the second area worth watching, since enterprise content runs across legal, technical, and multilingual document estates that vary widely in structure. Knowledge Catalog has to keep up with that heterogeneity before the context engine claim can be fully earned. Both areas are addressable given Google’s engineering depth, and the proof will land in production deployments over the coming year.
The transition from gen AI answering questions to agents acting on enterprise systems is, at its core, a context problem rather than a model problem. The platforms that will eventually scale agents from pilot into production are the ones that resolve business semantics, governance, and operational data at the platform layer instead of asking builders to assemble them per use case. Google has put more architectural weight behind that idea than any other major provider.
For technology leaders evaluating where to anchor their agent programs, three questions will define the next year of platform decisions:
Every major platform is converging on answering these questions from different angles, and the buyer’s job is to work out which architecture matches the data and governance reality of their own enterprise.
Mike Leone is a principal analyst at Moor Insights & Strategy covering data platforms and analytics, data infrastructure and storage, and data governance and enterprise data strategy. He brings 15 years of analyst experience from his work at Enterprise Strategy Group, where he rose to practice director for data management, analytics, and AI. Mike's work is grounded in a strong technical and strategic foundation, including early roles in software and hardware engineering.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。