At its Gemini at Work event this week, Google unveiled what it calls a single, universal agent: a prompt box that can plan a multi-step job, spin up temporary helper agents, run for hours or days, and deliver finished work back into inboxes, documents and code. Buried in the demo was a detail that tells you more about Google's ambition than any benchmark. When the agent decides which model should do a given task, its choices include not just Google's own Gemini but Anthropic's Claude, with more models promised. Google built an agent platform and then made its competitor's model a first-class citizen inside it.
That is not generosity. It is strategy. As independent analyst Carmi Levy put it, Google "isn't just looking to go toe-to-toe" on model quality; it is "trying to establish itself as the gatekeeper of the new enterprise operating system, powered by whatever underlying model makes the most sense." Thomas Kurian, Google Cloud's chief, framed it for customers as a shift in where work begins: "You give it objectives, not instructions. Work now starts in the prompt window." If the prompt window is Google's, it hardly matters whose model answers it. The best model for a task, Kurian noted, is not always the largest or the newest.
The pitch to nervous IT departments is control. Google is leaning hard on governance as its differentiator: every agent gets a cryptographically verified identity, its own audit trail, and permissions set by security administrators. Coworker agents receive company email addresses, calendars and storage, but see only the data their sharing rules allow. An Agent Sandbox isolates what they run; an Agent Gateway polices the network; spending limits pause an agent when it burns through its budget. In a market where OpenAI's always-on "dots" and Meta's "Muse" are chasing the same autonomous future, Levy argues Google is "arguably leading the conversation around identity, audit trails, and who, or what, is ultimately responsible for workflows."
The adoption numbers Google showed are striking, if self-reported. Nearly 500 enterprise customers have each processed more than a trillion tokens in the past year. Singapore's DBS Bank is chaining 70 to 80 specialised agents to draft corporate credit memos; Airwallex says Gemini cut its dispute-handling time by 40 percent. The same week, Spain's CaixaBank extended its Google Cloud alliance to 2033, betting a decade of its AI future on the platform. These figures come from the vendor and its customers, and the universal agent itself is still in private preview, so the right posture is interested skepticism rather than applause.
What nobody on stage could fully answer is the governance question the technology itself raises. As Info-Tech's Mahmoud Ramin observes, a basic chatbot keeps risk "much more contained," while an autonomous agent with access to many business systems, able to talk to other agents, is a different animal entirely. Enterprises will have to spell out which decisions an agent may make, what it may touch, and how its work is checked. Levy's own caution is the sharpest line of the week: Google's agent can write an email and build a slide deck like everyone else's, but whether a CIO can trust it "to run mission critical business processes indefinitely without doing something stupid enough to generate damaging headlines is another story altogether." The agent layer may be the ultimate prize of the AI era. Winning it means convincing buyers to hand a machine the keys, and promise it will not crash the car.