Agent-Ready Data Architecture: Streaming Real-Time Data to AI Agents

Watch the full session · 39 min

In a Striim-sponsored webinar, Striim and Google Cloud experts explained why data warehouses built for weekly reports leave AI agents with stale context, and how change data capture, event stream processing and MCP feed agents live operational data on Google Cloud.

AI agents need live operational data, but most enterprise data platforms were built to feed weekly reports. That gap was the focus of "Building Agent Ready Data Architectures on Google Cloud," a Data Science Connect webinar sponsored by Striim. Jake Bingston, VP of AI Solutions at Striim, and Schneider Larby, senior manager in Google Cloud's global partner technical architecture team, explained how streaming data into Google Cloud gives agents the fresh context they need.

Brian Mink, co-founder and president of Data Science Connect, framed the problem in his opening: "AI agents are only as effective as the data they can access."

Why are data warehouses not enough for AI agents?

Bingston traced the issue to how warehouses were designed. They were built for reporting, he said. A weekly report needed a week of data, not data from the last second, so teams built large batch jobs that loaded data daily or even weekly.

That works for dashboards that show a static point in time. Agents are different, according to Bingston, because they make operational decisions and need to know what is happening right now. Teams that prototype an agent on the existing warehouse often hit the same realization in production: "Oh actually this data is stale."

Connecting agents straight to operational databases is usually off the table, he added. Operational teams won't give that access to nondeterministic systems when no one knows what or how many queries an agent will run. The result, Bingston said, is a shift toward real-time operational lakehouses that agents can consume from, sometimes owned by platform and application teams rather than data teams.

What are the stages of real-time data maturity?

Bingston described three stages Striim sees organizations move through in support of their enterprise AI strategies:

1. Cloud migration and adoption: a lift and shift that puts data in the cloud, where many AI tools already run.
2. Data and platform modernization: unifying data in a modern architecture such as a lakehouse, loaded in real time rather than in batch.
3. Event stream processing: analyzing data in memory while it moves, identifying events of interest and passing them to a messaging layer such as Google Pub/Sub, where an agent picks them up and acts.

That third stage is what lets agents act autonomously instead of waiting for a manual trigger, Bingston said. He pointed to change data capture (CDC) from large operational databases such as Oracle and SQL Server as one of Striim's core strengths, along with enrichment, filtering, aggregation and vector embeddings applied to data in flight.

What are the components of an AI agent?

Larby broke an agent into four parts. The brain is the AI model, Gemini or a third-party model. Tools let the agent fetch data and take actions. Orchestration gives it memory and state, which he said separates an enterprise agent from a basic chatbot that forgets the conversation within minutes. The runtime is the compute it runs on.

On Google Cloud, he said, teams can deploy a containerized agent to Vertex AI Agent Engine or Cloud Run, both fully managed, or to Google Kubernetes Engine (GKE) when they want to manage infrastructure themselves. GKE offers the most flexibility, including the ability to bring any model and choose GPU types, while the managed options handle scaling and include built-in tools such as a RAG engine.

Larby also walked through the Agent Development Kit, the Agent2Agent (A2A) protocol for communication between agents, Gemini Enterprise and a memory bank feature that carries context across sessions.

How does MCP connect agents to streaming data?

Larby explained the Model Context Protocol (MCP) as a client-server standard. An agent acts as the MCP host and client, and lightweight MCP servers expose data sources and tools to it. When a question needs data from an external system, the agent's client initiates a handshake with the server and the model knows which tools to use.

Striim supports MCP, he said, so an agent can work with data while it is still in transit. Google's MCP Toolbox for Databases provides prebuilt tools for Google Cloud data sources, which can be the targets Striim streams into.

"Without data AI is essentially useless," Larby said.

Why is enterprise data the differentiator for AI?

Bingston closed on competitive advantage. "Your data is your differentiator," he said, because "everyone has access relatively to the same models." The organizations that bring their own data in best will give agents the context to know how to act.

He described Striim's work with Google Cloud as ingesting and delivering data into BigQuery and Vertex AI, using Striim Copilot to set up pipelines in natural language, and applying agents to data in motion to flag anomalies, create vector embeddings and discover PII.

Bingston also cited a recent National Bureau of Economic Research survey of about 6,000 executives in which, he said, 89% reported no measurable productivity gains from AI yet. He read the same research as optimistic for the next three years, and argued that better data infrastructure for agents is what will change the picture.

Where to learn more about agent-ready data

Data Science Connect's free webinars continue in 2027, with 24 live sessions across four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. Enterprise data and AI leaders can also meet in person at The AI Enterprise Conference on Tuesday, August 31, 2027, at Pier Sixty in New York City.

Watch: Building Agent Ready Data Architectures on Google Cloud edited on YouTube
Event: Webinars · More coverage: Webinars insights