Google’s Gemini Agent: What the New “Universal Agent for Work” Can Do

The most interesting sentence in Google’s Gemini agent announcement isn’t about what the agent can do. It’s this one, from Google Cloud chief executive Thomas Kurian: “Gemini is the agent, and the model underneath it is a separate choice.”

For two years, the AI race has been framed as a contest between models. Google, more than anyone, built its pitch around Gemini the model. Now it is shipping a product that will happily route your work to a competitor’s model, starting with Anthropic’s Claude. That tells you where Google thinks the money and the lock-in will be. It’s the agent sitting in your inbox and documents, not whichever model tops this month’s leaderboard.

This piece draws on Kurian’s launch post on the Google Cloud blog, Google’s shorter announcement, and reporting from TechCrunch and Reworked, all read in October 2026. Only a limited group of companies can use the agent so far. Everything below describes what Google says it does, not what it has been seen doing.

What Google actually announced

Strip away the event language and the Gemini agent is three things Google used to sell separately, now behind one prompt box:

• A chat assistant that answers questions using your company’s own documents and data.
• An autonomous worker you give an objective rather than step-by-step instructions. Kurian’s example is the kind of task that normally eats an afternoon: setting up a meeting with a team in another region. That means finding the right people, checking calendars and starting the email thread.
• A coding agent, which Google pitches at developers and data teams.

The agent runs in Google’s cloud, not on your laptop. In Kurian’s words, “work that takes hours or days keeps running after you close your laptop, and it is still there when you come back.” It keeps one set of memory across devices, and for bigger jobs it can spin up temporary sub-agents and coordinate them.

Google says you’ll reach it from the web, iOS, Android, Windows, Mac and the command line, and from inside Google Workspace, Microsoft 365 and Slack. Supporting Microsoft 365 and Slack matters more than it looks. Most companies don’t run purely on Google tools. An agent that only worked in Gmail and Docs would be a hard sell to them.

The Gemini agent as a “coworker”

The part of the launch that will provoke the most debate is the coworker mode. Google says an agent can take on a lasting role with its own identity and its own Google Workspace account. Reworked reported that this includes its own Gmail address, calendar and file storage, and that its edits show up under its own name.

On paper, that’s a sensible answer to a real problem. When an AI assistant acts using your account, every email it sends and every document it edits looks like you did it. If something goes wrong, nobody can easily tell whether a person or an agent made the call. Giving the agent its own identity fixes that: the audit trail shows the agent’s name, not yours.

It also raises questions Google’s announcement doesn’t answer. Who is responsible when a coworker agent sends the wrong figures to a client? What do colleagues and customers need to be told when they’re emailing an agent rather than a person? Those are management and policy questions more than technical ones. Every company that turns this on will have to work through them. We cover the wider idea in our explainer on coworker agents.

Why letting Claude in matters

By default, the Gemini agent chooses a model for each job. Users can also pick one. According to Google and TechCrunch, Anthropic’s Claude models are the first outside option, with open and private models to follow.

Kurian’s post lists Google’s own models as four families. Argon handles frontier reasoning, Flash speed and volume, Omni generative media and Gemma lightweight open-weight work. So a simple request might go to a cheap Flash model while a hard analysis goes to Argon or Claude.

There are two ways to read the Claude decision. The charitable one is that Google is giving customers what they ask for. Many large companies already use more than one model provider and don’t want to pick a single winner. The strategic one is that Google wants to be the layer every agent request passes through, whichever model answers it. Both can be true. Either way, it’s a notable shift for a company that spent the last two years telling everyone its own model was the product.

How the Gemini agent connects to everything else

An agent is only as useful as the systems it can reach. Inside Workspace, Google lists Gmail, Drive, Docs, Sheets, Slides, Chat and Calendar. Beyond it, the list runs to Microsoft Office and Teams, Slack, Confluence, Jira, Git, Salesforce, ServiceNow, BigQuery, Databricks, Postgres, Snowflake and files on your desktop.

It also supports the Model Context Protocol, the open standard that lets AI tools plug into other software. In practice, that means a company can connect the agent to almost any internal system that has, or gets, an MCP server. Our guide to the Model Context Protocol explains how that works and why it has spread so quickly.

Google adds a library of “skills”, reusable instructions for multi-step tasks that teams can write and share. It also says the agent keeps four kinds of memory: session, semantic, procedural and episodic.

The controls, and why they matter more than the features

Giving one agent access to email, files, customer records and code is the kind of decision that keeps security teams up at night. Google clearly knows this, because a large part of Kurian’s post is about control rather than capability:

• Identity. “Every agent gets its own identity, cryptographically attested and governed like an employee, with least-privilege permissions.”
• Audit. Every action is logged and attributed to the agent that took it.
• Agent Sandbox for running agent work in a contained environment.
• Agent Gateway, which Google calls “an AI network firewall enforcing your organization’s policies in real time.” Its example rule: agents may not open documents classified “Need to Know”.

That last idea is the right one. Rather than trusting an agent to behave, the company sets limits it physically cannot cross. Whether it works depends on configuration. Most breaches happen because someone set permissions too broadly, not because the tooling was missing. Our article on whether AI agents are safe covers the questions worth asking before you give any agent that level of access.

What the Gemini agent costs, and who can use it

Google hasn’t published a price. Kurian’s post describes cost controls rather than costs: routing simple work to cheaper models, “Smart Routing”, and real-time spending caps. Reworked reported that the agent comes with Gemini Enterprise “at no additional charge for customers where Gemini Enterprise is available”. It also said per-project caps pause the agent until someone resumes it. I couldn’t find the no-extra-charge line on a Google page, so treat it as reported rather than confirmed.

Availability is limited. Kurian’s post says industry versions for financial services and legal work are “now in preview”, with government, healthcare and retail “coming soon”. Reworked reported that only a limited number of companies have access at this stage. There is no general release date for the agent itself.

And for individuals: not yet. TechCrunch reported that Google plans a consumer version after the business launch but gave no timeline.

How it compares

Google is late to this particular party. Microsoft and OpenAI already sell work agents, which we’ve covered in our explainers on Copilot Autopilot and ChatGPT Work. All three companies want to be where your workday starts.

What distinguishes Google’s version, at least on paper, is the combination of model choice, agent identities and a central gateway for policy. Reworked quoted analysts arguing that Google wants to be the gatekeeper for company AI agents, with control as the selling point. That seems right. No launch event can answer the real test. Will large companies trust it enough to move beyond pilots into everyday use?

What I’d watch for next

• A price and a release date. Until Google publishes both, it’s hard for most companies to plan around the Gemini agent.
• Real deployments. The preview customers in finance and law will show whether the controls hold up in heavily regulated work.
• Independent scrutiny of the controls. Agent Gateway and agent identities are the heart of Google’s pitch. Outside security researchers testing them, and publishing what they find, would tell companies far more than any keynote.
• The consumer version. If the coworker idea reaches personal Gmail accounts, the questions about identity and disclosure get much more personal.

Gemini agent: FAQs

Sources

• Google Cloud Blog, Thomas Kurian, “Welcome to Gemini at Work 2026: Introducing the Gemini agent” (8 October 2026)
• Google, “Gemini at Work” (8 October 2026)
• TechCrunch, “Google brings agentic AI to Gemini, starting with businesses” (8 October 2026)
• Reworked, “Google debuts its all-purpose Gemini agent for work” (October 2026)