Gemini is now the storefront
The name says Gemini. The architecture says model marketplace.
Google's new work agent sits across Workspace, Slack, Microsoft 365 and business systems such as Salesforce, ServiceNow and Snowflake. A user gives it an objective; the agent gathers context, plans the job, calls tools and returns finished work in the app where it belongs. For assignments that take longer, it can create temporary sub-agents and keep running in Google's cloud for hours or days.
But the agent does not promise to run every step on a Gemini model. Google's Smart Routing can select Gemini or Anthropic's Claude for each job, and the company says other models are coming. That makes “Gemini” less a model choice than a product layer above several model providers.
The concession is also the strategy
It is tempting to read Claude support as Google admitting its own models are not always best. In the narrow sense, that is exactly what dynamic routing means: there are jobs where the system expects another model to deliver a better result or better economics.
The more important point is that Google still owns the customer relationship. If an employee asks a Gemini agent to build a financial model and the router chooses Claude, the work remains inside Google's interface, permission system, billing controls and audit trail. Anthropic supplies inference; Google supplies the workplace.
That is a stronger enterprise position than insisting on model purity. Companies already use several models. Google's pitch is that they should stop making employees choose among them and let a governed router make the choice instead.
One memory, many places
The product is designed to carry the same organizational context across devices and apps. In Workspace it can operate inline in Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar. Google also says employees can reach it through Slack, Microsoft 365 or a command line.
A separate coworker mode gives an agent its own Workspace identity, email address and calendar. It sees only the information shared with it, while administrators get identity controls, permissions, sandboxing and audit records. That is the part enterprise buyers will examine more closely than the demo: a persistent agent with a name and inbox is useful only if its boundaries are legible.
The missing proof
Google has described a broad system, not a finished public product. The Gemini agent remains in private preview. There is no standalone price, no general release date and no independent evidence that Smart Routing consistently picks the right model.
Routing also complicates evaluation. A workflow may behave differently as models, policies and prices change behind the same agent name. Enterprise teams will need logs that record which model handled each step, what data it received, how much it cost and whether the result can be reproduced.
What changed today
The AI labs spent the last two years trying to make users loyal to a model. Google's enterprise strategy points the other way: make the model invisible, keep the agent and the control plane.
If that works, the winning product may not be the smartest model. It may be the layer trusted to decide when not to use its own.
Sources
- [1] Google Cloud Blog, “Gemini at Work 2026: Introducing Gemini agent” (Oct 8, 2026)Read source
- [2] Reuters via LA Post, “Google Cloud introduces Gemini agent for work as AI race heats up” (Oct 8, 2026)Read source
- [3] Unite.AI, “Google Cloud Unveils Gemini, Its Universal Agent for Work” (Oct 8, 2026)Read source
- [4] Yellow, “Gemini Agent Runs On Claude Too As Google Says Nearly 80% Of Clients Use Its AI” (Oct 8, 2026)Read source