Google Cloud Agents Now Query BigQuery Sans SQL

TL;DR

• Google Cloud shipped the Agent Development Kit (ADK) to build AI agents that query BigQuery databases using natural language instead of SQL — demonstrated with a Marketing Optimizer Agent pulling metrics directly.
• The kit combines Gemini’s multimodal video analysis with quantitative data queries, deployed on Cloud Run with a FastAPI backend for production use.
• Google positions this as a move beyond novelty chatbots toward enterprise tools that deliver business value, competing directly with OpenAI’s GPT agents and AWS Bedrock.
• Ana Esqueda and Martin Omander walked through the full code and architecture in a Google Cloud Tech video tutorial.

Google Cloud Ships ADK to Kill the SQL Bottleneck

Google Cloud introduced the Agent Development Kit this week, a framework designed to let developers build AI agents that interact with BigQuery databases through conversational queries — no SQL knowledge required. The company demonstrated the toolkit with a Marketing Optimizer Agent that pulls campaign metrics from BigQuery by interpreting plain-English requests, combining those numbers with Gemini’s ability to analyze video content for qualitative insights.

The ADK runs on Cloud Run with a FastAPI backend, targeting production deployment rather than prototype demos. Google Cloud Tech engineers Ana Esqueda and Martin Omander presented the full code walkthrough and system architecture in a video tutorial, showing how the agent bridges structured data analysis and multimodal AI capabilities in a single workflow.

The company framed the release as a deliberate shift away from what it called ‘cute’ chatbots — experimental interfaces that impress in demos but collapse under real business requirements. Instead, the ADK aims to deliver agents that marketing teams, analysts, and operations staff can actually use to make data-driven decisions without waiting on engineering resources.

Why ADK Matters for Enterprises Drowning in Data Silos

Here’s the thing: most marketing teams can’t write SQL. And most data teams don’t have bandwidth to field every ad-hoc query that bubbles up from campaign managers trying to figure out why last quarter’s YouTube spend tanked. The ADK attacks that friction head-on by letting non-technical users ask questions in plain language — ‘show me conversion rates by region for Q4’ — and get back structured answers pulled directly from BigQuery.

But Google didn’t stop at text-to-SQL translation, which honestly feels table stakes in 2026. The Marketing Optimizer Agent layers in Gemini’s multimodal chops, analyzing video creative alongside performance data. That means a campaign manager could theoretically ask the agent to correlate creative elements — say, the presence of a product demo in the first three seconds — with conversion metrics, surfacing insights that live in the gap between quantitative dashboards and qualitative creative review.

I’ve watched too many companies build janky internal tools that bolt LangChain onto a database and call it agentic AI. This feels different — tighter integration, serverless deployment, and Google’s bet that Gemini’s multimodal edge gives it something OpenAI‘s text-heavy agents can’t match. If you think of traditional BI tools as vending machines — you know exactly what button to press — the ADK is more like a bartender who remembers your order and suggests something new based on what’s fresh.

The competitive stakes are real. OpenAI’s been pushing GPT-based agents hard, and Anthropic‘s tooling has captured developer mindshare. Google’s counterpunch hinges on BigQuery’s scale and Gemini’s ability to process video, images, and structured data in a single query. AWS Bedrock and Azure’s AI agent offerings don’t have an equivalent to BigQuery’s data warehouse reach, which gives Google Cloud a structural advantage if — and it’s a big if — enterprises actually adopt ADK instead of rolling their own.

Who wins here? Marketing ops teams that burn cycles waiting for data pulls. Product managers who need to slice metrics six different ways before lunch. Analysts who want to prototype insights without spinning up a Jupyter notebook. Who loses? Custom dev shops that charge five figures to build bespoke SQL-to-Slack integrations, and data engineers who spend half their week answering ‘quick question’ Slack DMs.

Gemini Multimodal and BigQuery Create Google’s Moat

Google Cloud didn’t invent the idea of natural language database queries — LangChain made text-to-SQL a commodity feature over a year ago. What the ADK does is package that capability with Gemini’s multimodal processing and Cloud Run’s serverless infrastructure into a deployment-ready stack. The video tutorial shows the agent pulling campaign performance data from BigQuery, then analyzing video creative to identify patterns — say, whether ads featuring testimonials outperform product demos — without requiring separate tools or manual correlation.

That combination matters because marketing decisions increasingly hinge on both quantitative metrics and qualitative creative analysis. Traditional BI dashboards excel at the former but ignore the latter. The ADK’s architecture — FastAPI backend, Cloud Run deployment, Gemini doing the heavy lifting — attempts to collapse that workflow into a single conversational interface.

Google’s timing aligns with broader enterprise frustration over AI prototypes that never ship. The Serverless Expeditions series, which this tutorial extends, has focused on Cloud Run as the deployment layer for AI workloads specifically because serverless scales without ops overhead. Companies that experimented with LangChain in 2024 often hit a wall moving from notebook to production — ADK targets that gap by providing a reference architecture that handles auth, scaling, and integration out of the box.

The BigQuery integration is the real differentiator. Reportedly, BigQuery processes exabytes of data across Google Cloud’s customer base, and enterprises already running warehouses there don’t need to migrate data or spin up new infrastructure. The ADK plugs directly into existing schemas, which lowers adoption friction compared to tools that require data duplication or API wrappers.

Cloud Run and FastAPI Signal Production-First Design

Google built the ADK on Cloud Run and FastAPI for a reason — both technologies prioritize production readiness over experimental flexibility. Cloud Run handles autoscaling and cold starts, which means an agent can sit idle during off-hours and spin up instantly when a user fires a query. FastAPI provides async request handling and automatic API documentation, which matters when non-technical teams need to understand what endpoints do without digging through code.

The architecture choice also signals where Google thinks agentic AI is heading. Not toward sprawling autonomous systems that replace human decision-making, but toward focused tools that automate specific workflows — querying databases, summarizing video, correlating disparate data sources — and hand results back to humans. The Marketing Optimizer Agent doesn’t run campaigns autonomously; it surfaces insights faster than a human analyst could manually.

That’s a more conservative vision than the fully autonomous agents some AI labs pitch, but it’s also more plausible for enterprise adoption in 2026. Companies want AI that accelerates existing workflows, not black-box systems that make decisions without explanation. The ADK’s design — transparent queries, structured outputs, human-in-the-loop by default — reflects that reality.

What should we watch as this rolls out? First, whether enterprises actually deploy ADK agents beyond pilot projects — Google Cloud’s challenge has always been converting demos into production workloads. Second, how OpenAI and Anthropic respond — both companies have agent frameworks, but neither has BigQuery’s data gravity. Third, whether Google expands ADK beyond marketing use cases into finance, supply chain, or customer support, where the same pattern — natural language query plus multimodal analysis — could unlock similar value.

And finally, whether the ‘no SQL required’ promise holds up when queries get complex. Natural language works great for straightforward requests, but ambiguity creeps in fast when users ask multi-step questions or need joins across schemas. If the ADK can handle that gracefully, it’s a genuine productivity unlock. If it forces users back into SQL for anything non-trivial, it’s just a nicer frontend on the same old bottleneck.

Source: Google Cloud Tech YouTube