Google wants employees to stop writing prompts and start delegating outcomes. At its Gemini at Work event on October 8, Google Cloud CEO Thomas Kurian introduced the Gemini agent, a single assistant that takes a goal, plans the steps and returns finished work inside Gmail, Docs, Slack or Microsoft 365.

The first wave of workplace assistants answered questions, while the Gemini agent is built to complete tasks that can run for hours or days. That shift raises new questions about cost, control, and which product name appears on the invoice.

Let me dissect what the agent does in plain terms. Google describes it as an assistant that onboards itself the way a new hire would, learning an employee’s tools, colleagues and work history before it starts. Employees give it objectives rather than step-by-step instructions. For example, a manager can ask it to set up a meeting with the usual regional event leads next week without naming anyone. The agent works out who those people are from past chat threads, checks their calendars and starts the email thread.

The agent runs in Google’s cloud rather than on the employee’s laptop. That design gives it one memory across phone, desktop and browser, and it lets a long task keep running after the laptop is closed. Google lists four kinds of memory: the task at hand, accumulated knowledge about the business, learned procedures, and a record of past work.

Two capabilities matter most to enterprises, starting with the coworker agent. It receives its own company email address, calendar and Drive storage, so colleagues can mention it in a chat space or a document comment the same way they would a team member. The second capability is model choice, since the agent picks a model for each job, choosing today between Google’s Gemini models and Anthropic’s Claude models, with other models planned. A routine summary can run on a cheaper model while a complex financial analysis goes to a stronger one, which Google positions as a lever for both quality and cost.

Gemini Agent Vs Microsoft Copilot Autopilot

Microsoft reached the market two weeks earlier. On September 25, it rebuilt Copilot around Autopilot , a long-running agent that gets its own Entra identity, email address and memory inside a customer’s Microsoft 365 tenant. Microsoft governs these agents through Agent 365, its control plane for agent identity, access and cost.

Both companies now sell the same idea: an AI colleague with its own identity, governed by the controls IT already applies to employees. The key difference is reach. Autopilot works inside Microsoft 365, while Google built the Gemini agent to operate inside Microsoft 365, Slack, and Workspace. That design gives Google a route into companies that never moved off Microsoft’s productivity suite.

Google also brings a large installed base to the launch, and the company reports that nearly 90% of the Fortune 100 use Gemini Enterprise and nearly 80% of Google Cloud customers use its AI products. Those figures count companies rather than employees, so they reveal little about how many people use the tools every day.

What The Gemini Agent Does Not Solve Yet

The agent is not yet broadly available to enterprise customers. Google says it will reach Workspace customers on select Business and Enterprise plan soon . The company has not published pricing for agent usage, and that gap is important because an agent working for days consumes compute the entire time. Google added real-time spend caps that pause an agent when a project reaches its limit. The cap protects the budget by stopping the work, and finance teams still have no usage history to forecast against.

Governance remains the other open item for enterprise buyers. Google promises a verifiable identity for every agent, audit trails that attribute each action to the agent rather than a person and a gateway that enforces rules such as blocking access to documents marked confidential. None of these controls has run at scale outside early testers. Memory raises another concern for buyers: an agent that learns a team’s procedures over two years holds institutional knowledge, and Google has not explained how a customer would take that knowledge elsewhere.

One Brand, Too Many Products

The bigger source of confusion for buyers is the name. Google uses Gemini for its model family, its consumer app, its Gemini Enterprise subscription, the developer platform that replaced Vertex AI in April and now its universal work agent. Gemini Spark, the always-on personal agent Google unveiled in May, adds another layer. Spark requires a personal Google account with a Google AI Pro or Ultra subscription and does not support work or school accounts. An employee who uses Spark at home and the Gemini agent at the office is therefore using two products that share a first name but follow different rules.

The naming also works against Google’s own pitch. Google describes the agent and the model underneath it as separate choices, yet it named the agent after its model, which means a task assigned to the Gemini agent may run on Claude. Microsoft followed a similar path with Copilot, a brand that now spans consumer chat, coding, an agent builder and Autopilot.

The Gemini agent is a compelling step toward delegating real work to software, and its ability to operate inside Microsoft 365 makes it a credible alternative for mixed-vendor enterprises. For technology leaders evaluating it, the first question concerns scope. Which Gemini product does the contract cover: the app, the agent, Spark, or the developer platform? The second question concerns the cost of running agents. How is agent compute billed beyond the seat, and who is notified when a spend cap pauses a critical task?

Customers who get both answers before the private preview ends will be in a far stronger position to negotiate when Google sets the price.