Google Cloud Data Updates Bring AI Agents Closer to Production
Google Cloud’s October data push puts agents closer to production
Google Cloud’s October updates point to a broader shift: data products are being reshaped to give AI agents trusted context, access to operational systems, and ways to run across different environments. That’s more significant than a stack of feature announcements. The harder question is whether these pieces make agents dependable enough for real work; several of the performance claims are based on internal benchmarks, so they deserve that caveat.October 26–30
- Knowledge Catalog adds self-learning enrichment and Knowledge Studio preview. Its enrichment agent learns from user and agent interactions, along with playbooks, code repositories, and business documents, to identify missing or stale context. Teams can package the resulting context into reusable sets. The new Knowledge Studio gives context engineers and data stewards a conversational way to explore their data estate, inspect a visual context graph, and review changes before users or agents see them. Google says internal benchmarks found that grounding agents in Knowledge Catalog made them nearly four times more accurate, with up to 33% lower latency and up to 75% fewer tokens. Those are striking figures, though the announcement doesn’t explain the benchmark conditions. Read the launch blog.
- Knowledge Catalog reaches semantic models teams already rely on (GA). It can now ingest semantic models and metric definitions from Looker LookML, BigQuery Graph, dbt MetricFlow, Snowflake Semantic Views, Databricks Metric Views, and SAP Core Schema Notation. The catalog links those definitions to lineage, data quality, and access controls in one context graph. Gemini Enterprise, BigQuery conversational analytics, and custom agents can retrieve a metric’s trusted formula, verified join path, and backing table in a single context call, without moving data or rewriting business logic. For teams trying to keep agent answers aligned with established business definitions, that integration is the point. Read the launch blog.
October 19–23
- BigQuery adds column aliases in GoogleSQL and makes cross-region table copy generally available. Column aliases can now be used in
SELECTandWHEREclauses through a GA opt-in, reducing nested subqueries and easing migrations from Snowflake and Teradata. Cross-region table copying is also GA, aimed at regional replication and disaster recovery. Review the BigQuery release notes. - BigQuery conversational analytics adds Canadian regions and CSV/PDF export. The service is available in Montréal (
northamerica-northeast1) and Toronto (northamerica-northeast2), addressing Canadian data residency requirements. Agent responses—including text, tables, charts, and citations—can now be exported directly to CSV or PDF (GA). Learn more about conversational analytics in BigQuery. - AlloyDB Omni 18.4.0 is GA, with query diagnostics and external key management. The release brings PostgreSQL 18.4 to self-hosted and multi-cloud environments. Its
google_insights_v2extension provides granular query execution and wait-event diagnostics. For Transparent Data Encryption (TDE), it adds Thales CipherTrust Manager support alongside HashiCorp Vault. Explore the AlloyDB Omni release notes. - Cloud SQL for PostgreSQL adds managed connection pooling and changes its private IP default. Existing Private Service Connect (PSC) instances can enable built-in connection pooling to support high-concurrency workloads. PSC is now the default private IP option for new Cloud SQL instances in new projects, simplifying cross-project networking. Review Cloud SQL updates.
- Unico says its Google Cloud data architecture has helped prevent more than $5 billion in fraud. The digital identity platform moved from Databricks Spark/Scala clusters to BigQuery and Spanner. By bringing real-time transaction streams together with petabyte-scale analytics, Unico says it can perform low-latency identity verification and has prevented over $5 billion in fraud. Read the Unico case study.
- A new guide shows developers how to start building Looker Data Apps. After the feature’s preview debut, the Google Developer Community published a walkthrough with practical patterns for creating prompt-driven applications inside Looker. Read the guide on the Google Developer Community.
October 12–16
- Looker Data Apps enter preview, turning reports into interactive applications. Developers can generate these web apps with natural-language prompts, the Antigravity CLI, or Claude Code, then host them inside Looker. Users can test scenarios, change parameters, and trigger workflows in downstream platforms. The apps inherit LookML semantic modeling, row-level security, and access controls, and incur no runtime LLM token costs. The promise is to move business reporting toward operational action; how far teams take that will depend on the workflows they connect. Read the announcement blog.
- The Data Agent Kit gets an end-to-end data science walkthrough. Following the kit’s general availability, a new guide demonstrates how data teams can build, test, and schedule machine learning pipelines from their preferred code editors. Data Agent Kit; Explore the walkthrough and codelab.
- LookML dashboard agents can now be managed as code. Teams can keep agent parameters, configurations, and prompt instructions alongside dashboard files. That brings Git version control, peer review, branching, and automated CI/CD validation to conversational BI, so agent changes can be checked before reaching production users. Read the technical guide on the Google Developer Community.
- AlloyDB adds native hybrid search with Reciprocal Rank Fusion. AlloyDB for PostgreSQL can combine vector similarity with keyword BM25 retrieval in one database engine. Reciprocal Rank Fusion (RRF) merges lexical scores and semantic embeddings into a ranked result without requiring a complex SQL query. The aim is to avoid operating and synchronizing a separate search engine alongside an operational database for retrieval-augmented generation (RAG) and search applications. Explore hybrid search in AlloyDB.
October 5–9
- Data Agent Kit becomes generally available, connecting coding agents to Google Data Cloud. The free kit includes Model Context Protocol (MCP) tools and Google-authored agent skills for more than 15 Google Data Cloud services. It works with coding agents in VS Code, Antigravity, Cursor, Claude Code, and Codex, letting developers and data practitioners inspect schemas, write queries, and build data pipelines in natural language from an IDE or CLI. Data Agent Kit; Read the announcement blog.
- Spanner Omni reaches GA for workloads deployed across environments. It brings Spanner’s distributed SQL, graph, vector, and key-value capabilities to on-premises data centers, other public clouds through Kubernetes or VMs, and local developer laptops. Organizations can use consistent APIs and query engines across hybrid and multi-cloud deployments, then scale to managed Spanner on Google Cloud. Read the launch blog; Spanner has also removed cumulative mutation limits for DML transactions.
- AlloyDB’s agent features span database architecture and search. Google announced PostgreSQL for agents in AlloyDB, built around an agentic database architecture designed for real-time data access at agent scale with workload isolation. Native BM25 ranking is in preview for AlloyDB and Cloud SQL, allowing developers to combine keyword search and vector embeddings inside PostgreSQL without maintaining an external search engine. For hybrid deployments, the AlloyDB Omni RPM Orchestrator is generally available. Google has also published a Build AI apps and agents with AlloyDB eBook.
- Lakehouse runtime catalog expands its regional coverage and Iceberg support. Regional endpoints are live in 22 Google Cloud regions across the Americas, Europe, Asia-Pacific, and the Middle East. They support both the Apache Iceberg REST catalog endpoint and the Apache Hive catalog endpoint, helping customers meet data residency and sovereignty requirements. Lakehouse runtime catalog regional endpoints. Flexible column names for Apache Iceberg tables are also generally available by default across external and managed tables, allowing queries against names with spaces, hyphens, and Unicode characters. Details on flexible column names. Google also announced a preview of cross-cloud caching and connections for the borderless Lakehouse.
Analytics and infrastructure updates
Three BigQuery additions target different parts of the analytics workflow. Six augmented analytics table-valued functions automate metric diagnostics, trend analysis and pattern discovery where the data is stored. They can also serve as skills for AI agents. TabFM brings tabular foundation model capabilities to predictive analytics, while identity columns generate sequential surrogate keys automatically. The latter is a smaller change, but it addresses a common source of friction in warehouse pipelines.
For larger AI workloads, Dataflow adds Pause/Resume controls and support for NVIDIA RTX PRO 6000 Blackwell GPUs, aimed at high-throughput streaming and batch pipelines. Spark teams can also use flexible VMs to reduce compute stockouts and improve availability.
Google reports gains across its data cloud customer base, though the examples span different products and workloads. Yahoo says flexible VMs in Managed Service for Apache Spark cut provisioning failures by 85%; PayPal moved mission-critical analytics workloads to the service to scale while reducing operational overhead. Surescripts moved 30.5 billion annual healthcare transactions to AlloyDB and reports 99.998% uptime. Airwallex runs more than 5,000 databases on Cloud SQL with 99.99% availability and a team of four DBAs. Lucius AI operates a global platform covering 210,000 tenders on AlloyDB and MCP, with ScaNN vector search reported to be 47 times faster.
More on Yahoo, PayPal, Surescripts, Airwallex and Lucius AI.
September 7–10
Pub/Sub can now run AI inference on incoming event streams using models hosted on Gemini Enterprise Agent Platform. The generally available AI Inference SMTs append each model prediction to its event, making it available to downstream systems such as BigQuery or Bigtable. That could simplify anomaly detection workflows, although the announcement doesn’t specify which models or inference patterns are supported.
Managed Service for Apache Kafka’s PostgreSQL source connector is also generally available. It captures database changes and sends them to Kafka with low latency. Supported sources include Cloud SQL for PostgreSQL, AlloyDB and self-managed PostgreSQL. It joins managed connectors for MirrorMaker 2.0, BigQuery, Cloud Storage and Pub/Sub. Questions or feedback can go to kafka-hotline@google.com.
Dataflow pause-on-failure is now generally available for batch jobs. It preserves a job’s state before failure, giving teams time to resolve external issues and resume without discarding completed work. The feature is intended to help manage costs and reliability during temporary outages or capacity constraints.
BigQuery has renamed the insertAll streaming API as the BigQuery Storage Write API (REST). Existing users need to make no code changes: the endpoint remains fully backward compatible. The stateless JSON-over-HTTPS API offers a lighter option than gRPC libraries for serverless web apps, IoT telemetry and AI logging. For high-throughput, continuous pipelines, Google still recommends the Storage Write API (gRPC).
August 31–September 4
BigQuery continuous queries now support stateful operations in Preview. Teams can use joins, aggregations and window functions in streaming queries—for example, to calculate a rolling 30-minute average and send richer real-time signals to applications or AI agents. The documentation has details; feedback is welcome at bq-continuous-queries-feedback@google.com.
Kafka users can now generate mock data for a cluster with a generally available synthetic data tool. It takes three clicks and can start streaming data in less than two minutes, without modifying a client application or setting up a virtual machine. For a quick cluster check or feature trial, that’s a useful shortcut. See the quickstart.
Dataflow pipeline updates are generally available with more flexible migration options. Teams can stop and replace a pipeline, or use a parallel pipeline to speed the transition from the old version to the new one and reduce disruption. A timeout on drains can also limit costs if processing gets stuck. See the upgrade guide and the instructions for updating a pipeline.
July 6–10
Google-managed Apache Iceberg tables for Lakehouse are in Preview and available in the console. The pitch is a shared storage layer that supports native reads and writes across multiple engines, avoiding duplicate pipelines and synchronization work between BigQuery and open-source tools. Teams can run concurrent DML and DDL operations across analytics tools, while automated management handles compaction and partition tuning. Whether that reduces operational work in practice will depend on how well the managed tables fit each team’s existing engines and workflows. Details are in the Lakehouse tables documentation.
June events
The “Beyond the Query: Powering AI Agents with Bigtable, Firestore & Memorystore” series covers updates to Google Cloud’s NoSQL database portfolio. It’s aimed at both newcomers and existing users interested in capabilities announced at Next ’26. Register for the series.
Google Cloud Labs is also running technical Cloud Engineer’s AI Toolkit workshops for platform engineers, SREs and cloud infrastructure teams working to take AI prototypes into production. The sessions cover data-driven work with BigQuery, AlloyDB, Gemini and other tools; more locations are planned. Toronto’s Data Cloud workshop is June 25, with registration here. Chicago’s Data Cloud workshop is June 30; register here.