Google's universal agent wants to do your enterprise work

Google wants you to close your laptop, walk away, and let AI keep doing your job until it is done.

At its Gemini at Work 2026 event, Google Cloud introduced the Gemini agent, an ambitious effort to turn generative AI from a scattered collection of chat windows into a single, autonomous workforce for enterprise operations.

According to Google Cloud CEO Thomas Kurian, the sheer volume of enterprise AI usage has crossed a tipping point. Over the past year, nearly 500 Google Cloud customers processed more than one trillion tokens each. Almost 80% of all Google Cloud clients now use its AI tools, and nearly 90% of Fortune 100 enterprises run on Gemini Enterprise.

A coworker with its own corporate email

Instead of making workers prompt a bot line by line, the Gemini agent operates on high-level goals. You hand it an objective, and it orchestrates the tools, creates documents, and executes code across Google Workspace apps like Gmail, Docs, Sheets, and Drive, as well as third-party suites like Microsoft 365 and Slack.

Because the agent runs persistently in the cloud, it remembers everything across devices without requiring users to repeat background context. If you shut down your computer in the middle of a massive data reconciliation, the agent keeps running.

It can also spawn temporary sub-agents to tackle multi-step projects in parallel, or take on a permanent identity. Google calls these coworker agents. They can receive their own dedicated email addresses, such as an @agents.company.com handle, along with scoped data access, persistent storage, and defined departmental responsibilities like a project manager or financial analyst.

To make this stick, Google gave the agent four distinct memory layers: session memory for active multi-day tasks, semantic memory for structured domain knowledge, procedural memory for learned routines, and episodic memory to recall previous actions. It also supports the open Model Context Protocol (MCP) and integrates natively with platforms like Jira, Git, Salesforce, Snowflake, and BigQuery.

Running Claude inside a Google ecosystem

Perhaps the most surprising technical decision is how Google handles underlying models. Gemini agent separates the execution layer from the raw model weights, routing tasks across Google's Gemini models and Anthropic's Claude, with plans to add more external models later.

Simple tasks get funneled to cheaper, lighter models, while complex legal or analytical tasks go to frontier models. Google is pairing this with Smart Routing and real-time spend caps to prevent budget blowouts. Customers are already leaning heavily into this approach: PayPal routes 10 million multi-model requests weekly, while Brazilian financial conglomerate Bradesco slashed document review times from one hour to five minutes and reduced risk inconsistencies by 60%.

Why it matters

We think Google made a brilliant strategic concession by baking Anthropic's Claude directly into its enterprise orchestration. Rather than insisting that customers rely exclusively on Gemini models, Google Cloud is positioning itself as the indispensable operating layer for all corporate AI. It recognizes that frontier model leadership changes every few months, but if Google owns the memory graph, the enterprise connectors, and the security policies, it wins regardless of which model processes the tokens.

For knowledge workers and enterprise teams, assigning autonomous tasks to an agent with an actual company email address transforms AI from a toy into an actual assistant. It removes the friction of copy-pasting between disconnected web tabs and solves the problem of AI losing its memory the moment a browser tab closes.

However, the risks are substantial. Granting persistent, headless agents deep access to internal databases, Slack channels, and executive inboxes creates a massive attack surface. If an autonomous coworker agent is tricked or misconfigured, the blast radius across connected enterprise platforms could be disastrous. Furthermore, coordinating multi-agent loops that run uninterrupted for days can rapidly rack up unpredictable API costs, even with Google's promised spend caps.