Google turns Gemini into an agent that works like a co-worker

Google announced on Thursday that Gemini will become a unified agent that can act on users' behalf — not just answer questions — and it starts with businesses, not consumers.

The company made the announcement at a Google Cloud event. The agent runs from a single interface and can plan work, generate code, schedule meetings, and connect to corporate systems. Google CEO Sundar Pichai framed the opportunity in numbers: Gemini today has over 1 billion monthly active users, and nearly 90% of Fortune 100 businesses use Gemini Enterprise at work.

The stakes are straightforward. AI tools are shifting from conversational chatbots to systems that take ownership of assigned tasks, and Google is entering a field where consumer-facing agents — Meta's Muse, messaging-based agents like Instinct, and ChatGPT's recently launched Dots — already compete. Google's distribution inside corporations is its differentiator. The catch: enterprise adoption is a claim about seats, not about how many of those seats will trust an agent to act.

Why businesses first?

Google will roll out the agent to companies before consumers. Pichai said the enterprise-first path lets the company solve the "harder problems around security, scale, and performance," which come with launching powerful agents.

That sequencing is a hedge as much as a strategy. Security and scale are genuine constraints, but launching to businesses also means Google can price, meter, and audit the agent before it faces the unpredictability of consumer use.

Takeaway: Google is testing its riskiest AI product on paying corporate customers first.

What can the agent actually do?

Google Cloud CEO Thomas Kurian said the new agent can be given "objectives, not just instructions." In practice, that means it:

• Plans the work required to hit a goal
• Uses custom skills and tools
• Connects to internal business systems
• Writes an audit trail attributed to the agent, not a person

By default, the AI picks the best model for the task. Users can override that choice and select a model themselves — including third-party ones, starting with Anthropic's Claude models. Google said the model picker will expand to open source models and other private models later. Notably, Google did not say which of its own models the default router prefers, or what the routing costs.

Requests can include attachments such as files, folders, or projects built for specific workstreams. The agent connects to business data and systems including Google Workspace, Microsoft 365, Slack, Jira, Confluence, Git, BigQuery, Databricks, Postgres, and Snowflake. It also works securely with any Model Context Protocol (MCP) server — an open standard for connecting AI models to external tools and data — inside or outside the company's network.

Takeaway: the pitch is breadth of integration, and Google named the integrations; it stayed quiet on pricing per action.

How do you supervise it?

Users track the agent through a "tasks inbox" interface. There they can see Gemini's thinking process, how it delegates tasks to subagents, which special skills and code it loads, and its progress on the job.

The agent also gets its own Workspace account, as if it were another employee. It has its own email address and its own context, Google said. It knows who sits on which team, their time zones, who approves items, and what is on people's calendars. Users reach it by tagging it, emailing it, sharing with it, or adding it to a group chat.

Giving an agent an identity cuts both ways. It makes attribution and accountability cleaner — every action carries the agent's name in the audit trail. It also means an AI with an email address and organizational awareness is only as trustworthy as the permissions granted to it. Google described the controls it offers for watching the agent work; it said less about how organizations limit what it can touch.

Takeaway: the tasks inbox is Google's answer to the control problem — visibility instead of restriction.

Where is it available, and who is testing it?

The agent will be accessible on iOS and Android mobile devices, Windows and Mac desktops, the command line interface, Google Workspace, Microsoft 365, ServiceNow, and Slack.

Early testers include the sportswear brand On, Shopify, and PayPal. Google also named Gemini Enterprise customers such as BNP Paribas, Bradesco, Merck, Orange Spain, Santee Cooper, SOMPO, Ulta Beauty, and Wesfarmers.

A tester list of three companies is thin evidence for a product of this ambition. The named enterprise customers use Gemini Enterprise broadly; Google did not specify which of them run the agent itself.

Takeaway: the rollout is real but early, and Google's customer names cover the platform, not necessarily this product.

What about cost control?

Google said new flexible spending options will help businesses keep enterprise AI costs under control. Those include:

• Multi-model orchestration (coordinating several models on one task)
• Smart routing
• Real-time spend caps

Cost is the quiet variable in every agentic AI deployment. An agent that plans, delegates, and calls multiple models per objective can consume far more compute than a chatbot answering one question. Google is addressing that risk head-on with caps — a sign it expects customers to hit spending limits, not a sign the agent runs cheap.

Takeaway: spend caps are a concession that agents are expensive by default.

What happens next?

The consumer rollout follows the business launch, on a timeline Google did not specify. The company also committed to expanding the model picker to open source and other private models.

The open question is adoption depth. A billion monthly active users and near-universal Fortune 100 penetration guarantee distribution, not delegation. Whether employees actually hand objectives — rather than questions — to an AI co-worker with its own email address is the test Google has now set for itself.