On October 8, Google Cloud unveiled something at its yearly event that is more than another chat box: the Gemini agent. It is pitched as a universal agent for work. You hand it an objective, it breaks the task apart, picks models, crosses applications, finishes the job, and comes back with the result.
The most counterintuitive part is that it keeps working after you close the laptop. The agent lives persistently in the cloud, shares one set of memories and a personalization graph across devices, and can run a job for hours or even days without stopping. It can also spin up coworker agents with their own corporate email addresses, independent storage, and workspace accounts, so the AI shows up inside the org as a managed entity rather than riding under your user identity.
Under the hood this is an orchestration layer, not a single model. Simple tasks route to a lightweight model, long-horizon heavy work runs on a high-end model, and today it can already call the rival Claude, with more private and open models planned. Cost controls include smart routing and real-time per-project spend caps that pause an agent the moment a budget threshold trips. The governance story is unusually enterprise-shaped: cryptographically attested agent identity, action-level audit trails, and a so-called agent gateway that Google openly calls an AI network firewall.
The tools registry already lists customer relationship management, ticketing, code platforms, data warehouses, office suites, and any protocol server. Customers are already in: a sportswear brand tested dynamic model selection, an e-commerce platform blends frontier models for millions of merchants, a payments company routes ten million multi-model requests every week, a Brazilian bank cut document review from one hour to five minutes, and a Spanish telecom has deployed more than a thousand custom agents.
The executive in charge threw in the muscle-flex numbers: nearly 90 percent of the Fortune 100 use this enterprise offering, and nearly five hundred cloud customers each processed more than a trillion tokens in the past year. The subtext is plain. Google has spent well over a hundred billion dollars on AI infrastructure and now urgently needs high-margin enterprise contracts to justify that bill.
But do not let the one-front-door pitch dazzle you. The agent is still in private preview, and most features have no published price or general-availability date. Handing an entire company's workflows to a cloud-persistent agent raises real questions about permission boundaries, audit granularity, and who is accountable when it goes wrong. None of that shows up in a demo video.
The more interesting detail is commercial. A Google representative indicated that although coworker agents get their own workspace accounts, the company does not intend to charge a per-seat license when the agent acts as a coworker, with pricing shared closer to wider availability. That shifts the cost model from selling seats to metering activity, which lowers the barrier to deploying coworker agents at scale once it works, but makes the bill harder to predict and the overage rules fuzzier.
Pull back to the competitive picture and it gets more interesting. Microsoft and the rival lab are building similar agent front doors, but Google's move is the most aggressive: it folds the office suite, email, and code platform into one persistent agent. The upside is fewer tool switches for users. The downside is that the lifeline of every workflow now runs through one orchestration chain. If that service goes down or its billing rules change abruptly, your whole office stalls, not just one plugin. For enterprises, convenience and single-point dependency are two sides of one coin, and the exit plan needs to be clear before any purchase.
One more layer almost nobody mentions: data sovereignty. This agent ships with a data layer that can query rival cloud storage directly, across clouds, with no variable egress fee. That sounds convenient, but it lays the path of data movement fully under its orchestration layer. For enterprises it means your data no longer just sits in one cloud. An agent now shuttles it across multiple clouds on your behalf. The compliance team will sooner or later have to answer one question: when the agent moves data across clouds for you, where exactly does the line of accountability fall. The other side of convenience is an audit surface quietly enlarged.
What this means for you: if you lead a team or manage knowledge work, Google's direction is worth watching, because delegating objectives instead of step-by-step instructions is moving from demo to routine. But do not flip everything over in the first wave. Pilot with one low-risk workflow, get permissions and audit trails solid first, then decide how much more of the business to hand over.