Google Collapsed Its Enterprise AI Lineup Into One Agent. It Also Routes Work to Anthropic's Claude.

Google Cloud CEO Thomas Kurian introduced what the company calls a "single, universal agent for work" at its Gemini at Work 2026 event on October 8, collapsing what had been several separately branded enterprise AI products into one. Users give it a goal rather than step-by-step instructions -- answer a question, build a slide deck, write and run code -- and it decides how to get there. Inside Gmail, Docs, Sheets, Slides, and Chat, it can simply be invoked with @Gemini. It is reachable from the command line, from Slack, and from Microsoft 365, and it is currently in private preview, with wide availability for Workspace Business and Enterprise plans promised only as coming "soon." Developers get their own entry point too: an API lets a company embed the agent as a "headless" worker with no chat window at all, wired directly into whatever internal tool is already handling the task.

For anything bigger than a single step, the agent creates its own temporary sub-agents, each with its own identity, running in parallel or in sequence over hours or days. A persistent version goes further: set up as a named "coworker" for a team or a role -- a finance analyst, say -- it can get its own Workspace account, its own email address, its own calendar, and its own Drive storage, with a presence in the company directory like any other employee. Google has not said whether every Gemini agent gets this by default, whether admins can toggle individual services, or whether it requires an extra license.

9to5Google's walkthrough gives a sense of what "give it a goal" looks like in practice: asked to set up a meeting with a regional team, the agent identifies attendees from a chat space and a past event thread, checks everyone's calendars, and starts an email thread to find a time -- including with people outside the company. Workspace Intelligence, a companion feature, can proactively flag a task as delegable in the first place: a manager's request for a slide deck in a chat thread gets a one-click handoff to the agent, rather than requiring anyone to think to invoke it. Underneath, Google describes four kinds of memory -- session, semantic, procedural, and episodic -- held together as one personalization graph per user, so an agent's sense of a person's preferences and history persists across every task it's handed, not just the one it's currently running.

The part that will unsettle Google's own cloud sales story more than anyone else's: the agent doesn't always run on Gemini. It routes each job to whichever model Google judges fits best, and Google says that currently includes Anthropic's Claude family alongside its own Gemini models, with support for other private and open models planned. A platform built to sell Google's own models is, by its own design, sometimes handing the actual work to a rival's.

Microsoft's own agent push -- hardware that ships Oct. 16 -- landed four days earlier with a different emphasis -- hardware and sandboxing over a single branded agent -- but the two companies are answering the same question: who gets to be the one layer an enterprise agent runs through. Google's answer leans on numbers it has not opened to outside audit. Kurian cited NTT Docomo saving 450,000 hours a year as a customer example; Pichai separately put overall adoption at "nearly 80 percent." Neither figure's methodology or sample was disclosed. Google's own Gemini 4 Argon, meanwhile, is said to score 77.9% on the DeepSWE v1.1 coding benchmark, ahead of Claude Opus 5.5, Claude Fable 5.1, and GPT-6 Astra -- a benchmark Google ran itself, on a day it was also announcing the product the score is meant to sell.

None of that is abstract. Collapsing several products into one agent, and letting that agent spin up identities of its own on demand, changes who has to actually manage the thing day to day -- and not everyone affected is Google.

Set the identity sprawl aside for a moment, though, and ask the more basic question: is any of this actually new engineering, or just a new coat of paint on products Google already sold separately?

Google has discontinued or folded together similarly named enterprise products before -- a GitHub-sourced dataset The Register cites puts the median lifespan of 308 discontinued Google products and services at about 4.1 years -- which is the quiet subtext of a launch promising, again, that this is finally the one agent to use. What's different this time is that the one agent openly admits it isn't always Google's own model doing the work -- a degree of candor that may matter more to a reader judging the claim than any adoption statistic Google is prepared to publish.