Agents moving toward safer database access

Cloud vendors are starting to let artificial intelligence agents touch live databases, but with tighter rules on what they can ask and return. (cloud.google.com) Google Cloud said on April 10 that its new QueryData tool is in preview and turns plain-language questions into database queries for AlloyDB, Cloud SQL, and Spanner. Google said the system uses “context sets,” a package of database-specific rules and metadata, to improve query accuracy. (cloud.google.com) Google’s product pages describe QueryData as a way to build “data agents” that can converse with a database instead of requiring a human to write Structured Query Language by hand. In release notes published April 6, Google said the preview adds “value search queries” that automatically match values and their context inside Cloud SQL and AlloyDB. (docs.cloud.google.com; docs.cloud.google.com) The pitch is simple: a database query is the instruction that pulls records from a table, and one bad instruction can expose the wrong data or produce the wrong answer. Google said QueryData is aimed at “regulated and mission-critical” use cases where companies need more predictable behavior than a general chatbot usually provides. (cloud.google.com; infoworld.com) That is the same opening Microsoft is chasing. TechCrunch reported on April 13 that Microsoft confirmed to The Information it is developing an OpenClaw-like agent for enterprise customers, with stronger security controls than the open-source tool it resembles. (techcrunch.com) OpenClaw’s appeal is that it can keep working on a user’s behalf, including multi-step tasks on a local machine, instead of waiting for one prompt at a time. The risk is that an always-on agent with broad access can make mistakes faster, which is why Microsoft’s reported project is being framed around enterprise controls rather than consumer convenience. (techcrunch.com) Google is making a similar tradeoff in a different layer of the stack. InfoWorld reported that analysts see QueryData’s deterministic, controls-based design as a way to make multi-agent systems more usable in sectors with compliance requirements, even if that adds setup work up front. (infoworld.com) Google tied QueryData to benchmark performance as well as customer deployment. The company said the tool builds on its top ranking in the BiRD text-to-SQL benchmark and said Hughes Network Systems has already put QueryData into production. (cloud.google.com) The pattern across both companies is that “agentic” software is moving out of demos and into systems that sit next to payroll records, customer databases, and internal documents. The selling point is no longer that an agent can act on its own, but that a company can decide exactly how far that autonomy goes. (cloud.google.com; techcrunch.com)