Google Launches Gemini Agent for Enterprise Work

Google Cloud turned its flagship AI assistant into an agent this week, launching a universal Gemini agent designed to do work across the business software enterprises already run. Announced October 8 at the company's Gemini at Work 2026 event in Mountain View, the move pushes Google's agentic ambitions from chat into persistent, cross-system task execution.

The launch lands in the middle of an enterprise agent land grab. Microsoft added a persistent Autopilot agent to Copilot in late September, and OpenAI rolled out always-on Dots agents days later. Google's answer is to position the Gemini agent as one layer that can operate across Google's own tools, Microsoft 365, Slack, and command-line interfaces — and, notably, to let it run on rival models.

A Gemini agent that picks its own models

The headline feature of the Gemini agent is dynamic model selection. By default, the Gemini agent picks the best model for each task, and users can override the choice — including selecting models from third parties. At launch, that means Anthropic's Claude models are supported, with Google saying it plans to expand the model picker to open-source models and other private models later.

The Gemini agent's architecture reflects a bet that model leadership will keep rotating. The agent keeps running in Google's cloud after the user closes a session, coordinates temporary sub-agents for multi-step jobs, and returns finished work inside the documents, inboxes, and developer environments where employees already work. Google Cloud CEO Thomas Kurian framed the philosophy as giving the agent objectives rather than instructions: it plans the work, uses custom skills and tools, connects to business systems, and brings back something finished, according to coverage of the keynote.

Early testers put those claims to the test across large deployments. Sportswear brand On tried the dynamic model-selection capability, according to reporting by Unite.AI, while Shopify blends frontier models for millions of merchants and PayPal routes 10 million multi-model requests per week. Brazilian bank Bradesco cut document review time from one hour to five minutes, and Orange Spain deployed more than 1,000 custom Gemini Enterprise agents.

Coworker identities, cost controls, and governance

Google is leaning hard on the enterprise controls that consumer agents skip. The Gemini agent can act as a coworker with its own dedicated identity — including agent email addresses and persistent storage — but its access is limited to specifically defined data and channels, with security, governance, and cost controls built in, according to CIO's coverage of the announcement.

Those guardrails matter because the new agent is meant to run for hours or days on long assignments. Spending caps can pause an agent once a project hits its limit, and every action lands in an audit trail. Google says it deliberately brought the agent to businesses first, before consumers, to work through what CEO Sundar Pichai described as the harder problems around security, scale, and performance, according to TechCrunch's report from the event.

The Gemini agent is available now in preview to a select set of customers, with wider availability coming soon, according to CIO Dive's reporting. Google has not disclosed a separate price or a general release date for the new agent.

What the scale numbers signal

Google used the event to show adoption momentum behind the platform the agent plugs into. Nearly 500 Google Cloud customers each processed more than one trillion tokens over the preceding year, nearly 80% of all Google Cloud customers are using its AI products, and nearly 90% of the Fortune 100 use Gemini Enterprise, according to Kurian's keynote figures cited by Unite.AI. Gemini itself now has over 1 billion monthly active users, Pichai said at the event.

The strategy is to become the orchestration layer rather than just another assistant. Instead of asking employees to come to Google's apps, the Gemini agent meets them where they already work — and can pick whichever model does the job best, even a competitor's. If enterprises accept that pitch, the moat is not the model at all but the permission layer, the audit trail, and the integrations sitting between the model and the work.

Watch for three things next: whether Google publishes independent benchmarks for the agent's cross-system task success, how quickly the model picker expands beyond Claude, and whether the preview pricing undercuts the enterprise agent bundles from Microsoft and OpenAI. The agent wars have officially moved from demos to deployment contracts.