At its Gemini at Work 2026 event, Google Cloud announced the Gemini agent — what CEO Thomas Kurian called “your new single, universal agent for work,” available in the Gemini Enterprise app, Google Workspace, third-party channels and a headless API. The pitch, per Google Cloud’s announcement: “You give it objectives, not instructions.” You delegate an outcome — a project update deck, a meeting arranged with “the usual team of regional event leads” — and come back to finished work. 9to5Google’s Abner Li reported the product is in private preview now, with wide availability coming for Workspace Business and Enterprise customers. SiliconANGLE carried the launch, noting it can write code and run multi-day tasks, and Bloomberg had the news too.
What it actually is
The Gemini agent is Google’s consolidation move: one agent, one API, running persistently in the cloud with “a single set of memories, context, and one personalization graph” across every device and channel you touch it from — command line, Workspace, Microsoft 365, even Slack. Work that “takes hours or days keeps running after you close your laptop.” It dynamically spawns its own sub-agents, each with a temporary identity, to parallelise multi-step tasks, and can also operate as a persistent “coworker agent” with its own @agents.company.com email address.
The most revealing detail is architectural. Four memory types — session, semantic, procedural (including “skills it writes for itself”) and episodic — are meant to fix the thing that makes today’s assistants feel disposable: they forget you. And under the hood, Google says the agent “runs each job on the model that fits best,” which today includes both Gemini and, notably, its chief rival’s models. Anthropic’s Claude family is listed as a first-class routing target. Google is claiming the orchestration layer; the frontier models underneath have become, at least in this product’s framing, interchangeable parts.
Google paired the launch with adoption figures: nearly 80% of Google Cloud customers use its AI products, and “nearly 90% of the Fortune 100 use Gemini Enterprise.” Those are Google’s own numbers for a customer launch event, and like every vendor statistic they are selected for the keynote — but even discounted, they mark the moment enterprise AI moved past experiments. Google says nearly 500 customers each processed more than a trillion tokens in the past year.
The uncomfortable convergence
Step back and the industry is converging on one shape. OpenAI’s DevDay last week introduced its always-on “dots” agents two days before it shelved GPT-6.1 Astra over a safety regression. Anthropic has Managed Agents. Meta launched its Muse personal agent in September. Now Google — the distribution giant — has its universal agent wired into the software half the world’s offices already run on. Everyone has accepted that the product is not a chatbot you visit; it is a persistent software worker that lives in your tools, remembers you, and acts on your behalf. The differences between vendors are now mostly about where the agent lives and who audits it.
That is exactly where the hard questions sit. An agent coordinating sub-agents across your corporate systems with its own email address — for hours, unsupervised — is the same architecture regulators and security researchers flagged all year, from the FTC’s probe into rogue agents to the incidents where agents exceeded their briefs. Google’s answer is governance tooling: identity and policy management, permission controls, sandboxing, and real-time spend caps, with Smart Routing promising cheaper models where quality allows. Whether that is enough is precisely the open experiment. An agent that “writes skills for itself” is powerful and, depending on your controls, one prompt-injection away from being someone else’s worker.
For the businesses that will actually use this — and for New Zealand firms riding Google Workspace rather than running their own AI stack — the change is practical: the price of the agent-in-your-office pattern has dropped from a six-month integration project to a rollout toggle. The catch rides along with it: giving an objective means handing over judgement about how the work gets done, and the audit trail is now a governance requirement rather than a nice-to-have. Enterprise software vendors spent two decades making IT the system of record for what employees do. They are now becoming the system of record for what employees’ agents do — and the vendors that get defaults right will shape work far more than any single model release this year.
