Google has unveiled a workplace agent built on its Gemini models that it describes as a universal assistant for work — software designed not just to answer questions, but to keep executing tasks in the cloud after the employee who assigned them has closed the laptop and gone home.
Announced October 8 as part of the company's Gemini at Work push, the agent can break a goal into steps, delegate pieces to specialised sub-agents, and let users choose which underlying model performs which part — including, notably, models from rival Anthropic. Work continues in persistent cloud sessions, so a research project, a data-cleaning job or a report assembly can run overnight and present results in the morning.
For Canadian businesses, the significance is less the launch-day spectacle than the direction of the entire industry. Every major platform vendor — Google, Microsoft, OpenAI, Anthropic — is converging on the same product shape: autonomous agents with memory, tool access and the ability to act across software systems. The competitive question has shifted from whose chatbot answers best to whose agent can be trusted with actual work.
That trust question is why enterprises are moving carefully. An agent that can act can also err at scale, touch data it should not, and spend money. Google's enterprise pitch leans on administrative controls, audit trails and the choice of models per task — letting a company route sensitive work to a model it trusts and cheap work to a faster one. The inclusion of a competitor's models inside Google's own agent platform is the clearest sign yet that customers, not vendors, will assemble these stacks.
Canadian technology leaders have seen this movie at smaller scale with each software wave. The practical advice from early adopters is unglamorous: start agents on bounded, checkable work — summaries, reconciliations, first drafts, monitoring — instrument everything, and expand authority as the error record earns it.
Google is betting that the winning workplace AI will not be a place employees visit, but a colleague that never logs off. Whether that prospect delights or unsettles depends on which side of the delegation you sit — but the products are no longer demos. They are shipping, priced, and pointed at the knowledge-work mainstream.
The multi-model detail deserves more attention than it will get. Google letting customers route work to Anthropic's models inside its own agent platform would have been unthinkable in the platform wars' first decade; in the agent era, it reads as the market imposing Switzerland. Enterprises do not want five agent subscriptions with five permission models — they want one control plane, and they are telling vendors to interoperate or be routed around.
The risks scale with the capability. Agents that act across systems inherit every security question a company asks about an employee — access, segregation of duties, auditability — plus one employees never pose: they can be manipulated by the data they read. Security teams are building the same defences they built for privileged users: least-privilege access, transaction limits, human approval above thresholds. The vendors that win the enterprise market will be the ones whose governance tooling matures as fast as their models.
For workers, the honest summary is transition, not replacement. The first agents absorb the work nobody applied to do — reconciliation, summarisation, monitoring, draft assembly. Job descriptions will quietly rewrite themselves around supervising that work. Gemini at Work is Google's claim on that future; the claims get tested, as always, one deployment at a time.
