A new tool from Google Cloud aims to improve the accuracy of AI agents querying databases in multi-agent systems or applications.

QueryData, which translates natural language into database queries with what the company claims is “near 100% accuracy,” is being pitched as an alternative to direct generation of queries by large language models (LLMs), which Google says can introduce inaccuracies due to their limited understanding of database schemas and their probabilistic reasoning.

However, to create that necessary understanding, enterprise teams using QueryData first need to define what Google describes as “context” describing how data should be accessed and queried, which involves encoding details about database schemas, including descriptions of tables, relationships, and business meaning, along with deterministic instructions that guide how queries are generated or executed.

Once the context and the guidelines are configured, teams can use the Context Engineering Assistant, a dedicated agent in Gemini CLI, to iteratively review the accuracy of queries against the Evalbench framework until they’re satisfied with the results.

After that, QueryData can be integrated into agent-driven workflows, where it acts as the execution layer between user requests and underlying databases.

It can be used within Google Cloud’s own data agents, currently available in BigQuery, or invoked via APIs by enterprises building custom agents and multi-agent systems. It currently supports AlloyDB, CloudSQL for MySQL, CloudSQL for PostgreSQL, and Spanner.

In custom setups, the agents handle reasoning and orchestration, while QueryData is responsible for generating, validating, and executing queries against data sources, returning results that can be used in downstream actions or decision-making, Google explained in a blog post.

Creates new workload category

The new tool, according to Pareekh Jain, principal analyst at Pareekh Consulting, marks a shift from tool-based AI to outcome-bound agents with built-in guardrails, which should help enterprises move multi-agentic systems and applications into production, and enable “decision-grade use cases” across finance, operations, and supply chain departments.

However, he cautioned, though QueryData reduces the need for prompt engineering for developers and improves reliability at runtime, it shifts the burden to upfront design and ongoing maintenance.

“It requires explicit schema understanding, deterministic instructions per data source, and ongoing maintenance as schemas evolve,” he pointed out. “This effectively creates a new workload category of data access engineering for agents.”

He said, “the tradeoff is clear. Without QueryData, systems are faster to build but unreliable in production, and with it, they are slower to build but viable at scale.”

This tradeoff, according to Jain, will ultimately influence enterprise usage patterns, with adoption likely to be strongest in regulated and mission-critical environments, while remaining slower in lightweight or experimental use cases.

Targets data layer as rivals bet on connectors, copilots

Further, Jain noted that the new tool also signals a broader strategic play by Google Cloud.

 “QueryData shows Google is trying to create a standard way for AI agents to safely access and use data. While OpenAI focuses on APIs, AWS on connectors, and Microsoft on apps like Copilot, Google is focusing on the data layer itself, on how agents actually talk to databases,” Jain said.

“This approach has strengths, especially with tight integration into Google BigQuery and Google’s data expertise. But it also has challenges, as it needs more upfront setup and is less flexible across platforms. Microsoft, in this case, seems to have an edge, because its tools are already built into everyday apps that people use,” he noted.

The risk for Google, Jain added, is that simpler approaches from AWS or Microsoft could confine QueryData to advanced use cases instead of making it a mainstream standard.

QueryData is currently in preview.