Google Gemini agent: One Work Bot That Also Runs Claude

The Google Gemini agent is Google Cloud's biggest swing yet at the "one AI coworker" idea. Announced on October 8 at Gemini at Work 2026, it's pitched as a single, universal agent that plans work, uses tools, taps company systems and hands back finished output inside Docs, Gmail and dev environments. It's also happy to route jobs to Claude models. Yes, really.

What actually happened

Google's blog says the agent has an organization's business context and handles everything from Q&A and knowledge work to content creation and coding. It "chooses the best model for the job," includes built-in cost controls, and ships with enterprise security, admin and governance features.

According to 9to5Google, it shows up in the Gemini Enterprise app, Workspace and third-party services. Google's example: your manager emails asking for a project update as a slide deck, Workspace spots that as a task you can delegate, and you hand it off with a single click.

Google also says the agent can act as a "team member," working for a whole group like a project manager would, or filling a specific role, such as an analyst in a finance department. That's a bigger promise than a chat sidebar.

Another example is scheduling. You ask for a meeting with "the usual team of regional event leads" without listing names. The agent works out who they are from your chat space and past threads, checks calendars and starts the email thread.

How the Google Gemini agent works

Under the hood, the agent "runs each job on the model that fits best." Google says that means Gemini and Claude models today, with other private and open models to come. Google pitches that routing as a way to "deliver optimal quality and lower your costs." It splits multi-step work across sub-agents that can run in parallel or in sequence.

Memory is the big flex. 9to5Google lists four memory types, including session memory for long tasks, semantic memory built from documents and conversations, and procedural memory for how jobs get done, including skills the agent writes for itself.

You can call @Gemini inside Gmail, Docs, Sheets, Slides and Chat. Google also lists the command line, Slack and Microsoft 365 as channels, plus an API that lets it run as a "headless" agent inside other apps.

Give it objectives, not instructions.

Who is already testing it

Google's press corner lists a stack of same-day customer announcements. Cooley is building a litigation agent, The RealReal expanded its Ask TRR shopping agent, and Zip US is building what it calls an AI-native product factory. Balyasny is using Gemini models for research, and Orange Spain says it will deploy more than 1,000 custom agents.

Those are Gemini Enterprise stories in general, so not all of them confirm the new agent specifically. Still, they show where Google is selling: regulated, document-heavy enterprises that already live in Workspace.

The catch

Private preview means most people can't touch it yet, and Google hasn't published a date or price in the posts we reviewed. Multi-model routing sounds great, but buyers will want to know which model handled which task, where data goes and how costs get billed.

There's also brand confusion. A "Gemini" agent running Claude jobs is a bit of a meme, and some 9to5Google commenters said so. It also lines up with our earlier coverage of Claude for Google Workspace, which already put Anthropic inside the same apps.

What to watch next

Three things will decide whether this is real or just keynote theater: the price, which Workspace plans get it, and how much control admins have over model routing and memory. Google says wide availability is coming soon for select Business and Enterprise plans, but it hasn't said when.

Also watch the API. A headless Google Gemini agent that other apps can call turns Google into an agent platform, not just a feature inside Gmail, and that's where the real enterprise money would be.

What to do as a reader

If your company runs on Workspace, ask your admin whether you're on a plan that's eligible for the Google Gemini agent preview. In the meantime, map one boring, repeatable task, like weekly reports, as a test case. Read the official Google Cloud announcement rather than demo clips.

For the model-race backdrop, see our notes on Gemini 4 Argon and the October AI agent odds market. Treat agent output like a junior colleague's draft: review it before it goes to a client. Nothing here is investment advice.