What we announced in streaming AI at Next ‘26

Every device, user, and microservice generates data. Ingesting this data, extracting meaning and insights, and driving business decisions in real time has the potential to deliver transformational business value.The rise of agentic AI represents an opportunity for users to overcome the challenges inherent in real-time analytics. But while agentic AI has the potential to accelerate adoption, users face a new set of challenges with effectively leveraging real-time data:

• Real-time context is hard to implement. Teams will choose to incorporate data from batch-oriented approaches, like periodic database syncs and scheduled refreshes. Agents have to either rely on stale data or require memory-intensive context windows. This “context lag” makes them ineffective for real-time agentic tasks like fraud detection, dynamic e-commerce recommendations, or autonomous supply chain adjustments. 

• Real-time systems are inflexible. Agentic tools lack the modularity to adapt to customer-specific requirements, forcing organizations to make difficult architectural choices. Data practitioners need a platform to meet them where they are, where they are free to make the tradeoff between latency, accuracy, and cost. 

Google Cloud provides a tightly integrated, unified streaming data platform that delivers both fully managed, Google Cloud-native services, as well as open-source-compatible services, and that come together to power large-scale AI training and inference. The platform is comprised of five key services: 

• Pub/Sub: Highly reliable, serverless, and fully managed service for messaging and event streaming that’s integrated with BigQuery, Dataflow, and Cloud Storage. Pub/Sub is utilized by organizations like Anthropic. 

• Dataflow: A serverless engine for batch, streaming, and now agentic AI. Leading enterprise organizations like Palo Alto Networks use Dataflow, as do Google services like Waymo and Google Maps. For instance, Waymo cars use Dataflow to help it “see” the world, plan their routes, and predict obstacles. Before a car hits the actual pavement, it “drives” millions of miles in a simulator, with Dataflow generating training datasets and validating the models that are used for autonomous driving.

• Managed Service for Apache Kafka: The fully managed way to run the popular open source streaming storage and data integration system on Google Cloud that’s highly reliable, secure, and cost efficient. Across the largest enterprises and startups, Apache Kafka serves as a staging location for critical training data and real time updates to AI agent context. 

• BigQuery: A unified platform for real-time ingestion and analysis. The Storage Write API provides high-throughput streaming into BigQuery and Lakehouse for Apache Iceberg tables with exactly-once delivery semantics and stream-level transactions. Additionally, BigQuery continuous queries enable real-time AI inference directly within the data pipeline by calling generative functions like AI.GENERATE_TEXT, allowing for immediate insights as data is ingested.

• Bigtable: Google’s NoSQL real-time database for processing streaming data from Pub/Sub and Dataflow automatically using  continuous materialized views, delivering results in seconds that are ready for low-latency serving using Bigtable’s in-memory tier.

Moving from insight to autonomous action

At Google Cloud Next, we announced a set of streaming AI capabilities to the Agentic Data Cloud, providing autonomous agents with instant context and enabling real-time actions, helping organizations feed real-time context to their AI agents.

For instance, imagine a supply chain agent that doesn’t just monitor IoT data, but autonomously reroutes a shipment around bad weather, confirms new delivery windows with the receiving warehouse, and updates the customer’s portal — all before a human supervisor is even aware of the problem. Consider a financial services agent that identifies a fraudulent transaction pattern, instantly freezes the account, communicates with the customer via their preferred channel, and initiates a new card shipment — all within seconds of the suspicious activity. Whether you’re creating embeddings on streaming data to power search, or building a sophisticated multi-agent fraud detection system, these new capabilities add powerful tools to your toolbox. 

Let’s take a closer look at these new capabilities. 

New streaming AI capabilities

At Next ‘26, we launched tightly integrated capabilities to our platform across three key areas:

1. Providing real-time, enriched context for agents

1.1. Pub/Sub AI Inference SMT (GA): You can now run inference on messages streamed through Pub/Sub. Data practitioners can choose any models available on Gemini Enterprise Agent Platform. Pub/Sub makes the inference call and appends the result to each message before sending it downstream, bringing Pub/Sub’s simplicity together with the Gemini Enterprise’s fully managed tools.

1.2. Pub/Sub Bigtable subscriptions (Preview): Stream Pub/Sub data directly to Bigtable. Pub/Sub Bigtable subscriptions directly materialize event data from a Pub/Sub topic into a Bigtable table, eliminating the need for custom pipelines and dramatically simplifying your streaming architecture. For instance, you can easily ingest vector embeddings into Bigtable to power semantic search workloads. 

1.3. BigQuery continuous queries stateful data processing (Preview): BigQuery continuous queries can now perform complex correlations between multiple data streams using JOINs and calculate metrics over consistent time intervals with tumbling window aggregations. This enables sophisticated analysis, such as calculating 30-minute averages or correlating events across different streams, directly as data is ingested into BigQuery. Furthermore, you can integrate AI directly into your data pipelines by calling generative functions like AI.GENERATE_TEXT, as well as materialize continuous query SQL results into BigQuery tables or export them to operational sinks like Bigtable, Spanner, and Pub/Sub for real-time reverse ETL.

2. Direct agents to manage your resources

2.1. Model Context Protocol (MCP) support for Pub/Sub, Managed service for Apache Kafka, Bigtable and BigQuery (GA): Your agents can manage Pub/Sub,Managed service for Apache Kafka services, and BigQuery using fully managed MCP endpoints. Agents can also publish messages to Pub/Sub. 

2.2. ADK integration (GA): Your agents can interact with your real-time data stored in Pub/Sub, Bigtable, BigQuery, or other Google Cloud services using pre-built ADK integrations. Developers can build agents acting on real-time context without having to implement complex configurations or plumbing.

3. Combine multi-agent systems with your data processing

3.1. Event-driven autonomous agents: As agents become core to our workflows, real-time data pipelines must evolve to incorporate them directly into the stream. We have enabled this capability by treating agentic logic as a first-class citizen within the Dataflow pipeline. You can now incorporate your agent code using the Agent Development Kit (ADK) and deploy it as a specialized node using the RunInference transform and the new ADKAgentModelHandler. Key advantages of this approach include:

• • Massive scalability: Leverage Dataflow’s architecture to process high velocity events upstream and keep hundreds of agents sessions active simultaneously, each driven by specific incoming events.
  • Pre-processing power: Dataflow handles the heavy lifting of complex data enrichment, delivering a “ready-to-act” context directly to the agent so it can focus on reasoning.

3.2. Dataflow Unified embeddings Sinks: We are introducing unified embedding generation directly within the data stream to eliminate “context lag”. You can now transform incoming data into high-dimensional vectors at low latency using Dataflow. These real-time embeddings are then seamlessly materialized into our expanded suite of high-throughput vector sinks, which now includes Cloud Spanner (featuring its new built-in vector search) and AlloyDB, providing you with an up to date vector database for semantic search needs as well as for your autonomous agents making RAG calls with an instantly searchable and perfectly synchronized long-term memory. This feature works with both remote and local models, for example Gemma. 

As we continue to build out the platform, customers can expect to see even tighter integrations and more powerful capabilities. We look forward to seeing what you build with these new capabilities.