Bloomberg Media’s data team just boosted SQL query accuracy by 63% by linking their knowledge agents to Google Cloud’s Knowledge Catalog. This is the first real-world win Google is flagging for its new Gemini Agent—a tool billed as a single, universal agent for enterprise work.
It’s a work delegate, not a chatbot. Unlike regular chatbots that wait for specific commands, Gemini Agent works on objectives you set. You tell it “Pull Q3 ad spend data, cross-check with campaign performance, and draft a summary for stakeholders” — it plans the steps, picks the right tools (like BigQuery for SQL, Docs for writing), connects to your company systems, and hands over finished work. It’s built to stay in the background: jobs that take hours or days keep running even if you close your laptop. It can also spin up temporary sub-agents for specific tasks, and runs on whatever model fits best right now. Today that means Google’s Gemini models and Anthropic’s Claude, with more private and open models planned. In Google Workspace, it even gets its own identity—an email address, calendar, and Drive entry—so colleagues can @mention it in Chat or Docs, and its edits show under its name in version history. For Google, this agent is the core product; the model behind it is just part of the routing decision, not the main offering.
Built for enterprise rules and budget control. Enterprises don’t just want tools that work—they want tools that are secure, trackable, and don’t blow IT budgets. Gemini Agent checks these boxes on key fronts. Governance is central: each agent has a unique, crypto-verified identity with minimum access rights, so it can’t peek at data it shouldn’t. Every action is logged, and agents run in a sandbox with a network firewall that enforces consistent rules across all use cases. Cost controls are another highlight: teams can set hard, per-project spend caps in the Cloud Billing Console. If the agent hits that cap, it pauses automatically—no surprise bills. Google says its TPU 8i system delivers 80% better price-performance than the previous generation, cutting operational costs. It also integrates with three of Google’s enterprise data services: Knowledge Catalog (maps business definitions once for all agents), Smart Storage (enriches unstructured data, which makes up 90% of enterprise data), and Borderless Lakehouse (queries Amazon S3 and Azure Data Lake without extra egress fees).
The gaps and open questions. Compared to competitors, Gemini Agent has clear edges. Microsoft 365 Copilot is tied exclusively to Microsoft’s ecosystem, with no multi-model support or persistent agent identity. Amazon’s QuickModels and OpenAI’s ChatGPT Work lack the same level of built-in governance and long-running job support. But there are big unknowns Google hasn’t resolved yet. No concrete benchmarks have been released—buyers have to evaluate the tool based on anecdotes like Bloomberg’s, not reproducible test data. There’s no general availability date, either, and public pricing is not available. That makes it hard for enterprises to do the due diligence they need before committing.
So far, Gemini Agent is betting that enterprises are tired of juggling multiple tools for different work tasks—one for coding, one for knowledge queries, one for content creation. A single agent that can handle all of that, with built-in governance and cost controls, is a compelling idea. But until Google sorts out the lack of benchmarks, GA date, and pricing, it’s still more of a promise than a practical solution for most companies. The question is: Will enterprises choose a universal agent that works across multiple models and tools, or stick with tools tightly integrated into their existing daily workflows?
素材来源：MarkTechPost · AI情报、大模型、AI智能体、AI应用落地
查看报道原文
本文由 WeValue 情报引擎基于公开报道整理改写
发表第一条评论吧