Google's Gemini Agent Keeps Working After You Log Off

Sundar Pichai, CEO of Google and Alphabet, who opened Google Cloud's Gemini at Work 2026 event. Photo: Lukasz Kobus / European Commission, CC BY 4.0, via Wikimedia Commons.

The most interesting hire at Google's big AI event this week wasn't a person. It was a coworker with its own email address, no body, and permission to keep working long after you've shut your laptop and gone home.

On Thursday, October 8, Google Cloud used its Gemini at Work 2026 event to drag Gemini into the agentic age. The headline: a single, universal AI agent for work that doesn't just answer questions but takes ownership of assignments — planning the work, calling tools, connecting to a company's internal systems, and returning finished results inside the documents, inboxes and developer environments people already use. And yes, Google says it will keep grinding away on multi-hour tasks while your laptop sits closed on your desk.

This is Google's biggest swing yet at the idea that the chatbot era is over and the coworker era has begun. The question is whether anyone actually wants an AI colleague that never clocks out.

What Google actually announced

Thomas Kurian, CEO of Google Cloud, laid out the agent in a keynote that the company later adapted into a detailed post on the Google Cloud Blog. The pitch, in Kurian's own framing, is that the agent can be given "objectives, not just instructions" — a line that will end up in a hundred enterprise slide decks by Monday.

In practice, that means the agent plans the work itself, selects skills and tools, connects to Cloud customers' business systems, and hands back finished work. Ask it to research a market, build a financial model in Sheets and turn the findings into a presentation — the agent is supposed to handle the whole chain rather than waiting for you to prompt each step.

The agent lives where work already happens: Gmail, Drive, Docs, Sheets, Slides, Chat and Calendar, and — notably — it also works through Microsoft 365 and Slack. Google isn't pretending your company runs entirely on Workspace. That's a small, honest concession, and it's the right one.

It also comes with the unglamorous plumbing enterprises actually care about: built-in cost controls, security, administration and governance. Agents that touch real business systems need real guardrails, and Google led with that.

It doesn't sleep when you do

The detail that made everyone sit up is the persistence. According to Google, the agent can keep working after an employee closes their laptop, letting assignments that take hours — or even days — run in the background.

We cross-checked Google's announcement against Reuters' and TechCrunch's same-day reporting, and the persistence claim, the customer names and the model-picker detail all line up across all three. This isn't a game of telephone; the facts survived three newsrooms.

Reuters also reported the strangest new primitive in the package: "coworker agents" that act as team members, with their own email address and access only to the information they're given. Your next colleague might have no chair, no lunch break and an @-address issued by IT. Google says versions built for financial services and legal work are already in preview, with government, healthcare and retail versions coming soon — sectors where a tireless, auditable worker is genuinely appealing and, frankly, a little unnerving.

The model-picker twist nobody expected

Here's the part that would have been unthinkable two years ago. By default, the agent picks the best model for each task — and at launch, that model picker includes not just Google's own Gemini models but Anthropic's Claude models too. Users can override the choice themselves. Google says open-source models and other private models will be added later.

Let that sink in. Google is shipping a flagship enterprise product that can route your work to its fiercest rival's model when that model is the better tool. It's the kind of pragmatism that only a company confident in its distribution can afford — or, less charitably, an admission that no single lab has the best model for everything anymore.

Early testers named in Kurian's post include sportswear brand On, which tried the dynamic model-selection capability, and Shopify, which blends frontier models for millions of merchants. PayPal is routing 10 million multi-model requests a week through the system. Brazilian bank Bradesco cut document review time from an hour to five minutes, and telecom operator Orange Spain has deployed more than 1,000 custom Gemini Enterprise agents.

The numbers Google wants you to see

Sundar Pichai, CEO of Google and Alphabet, opened the event with scale. Gemini, he said, now has over 1 billion monthly active users, and nearly 90% of Fortune 100 businesses use Gemini Enterprise at work.

Kurian added the enterprise depth: nearly 500 Google Cloud customers each processed more than one trillion tokens over the preceding year, and nearly 80% of all Google Cloud customers are using its AI products.

Pichai also explained why Google is starting with businesses rather than consumers: it lets the company solve the "harder problems around security, scale, and performance" that come with powerful agents, as TechCrunch reported from the event. Translation: agents that can touch your company's systems are dangerous, and Google would rather learn those lessons on corporate IT departments than on a billion consumers.

That caution is well-founded. This month alone we've covered tens of thousands of rogue AI-agent incidents and agents breaching government websites. An agent that keeps working after you close your laptop is a feature right up until it becomes a liability report.

The race context Google can't escape

Google didn't invent this category — it's playing catch-up in a race that accelerated hard this year. Reuters noted the announcement comes after OpenAI's September launch of always-on agents, reportedly called "dots," that chase user goals across apps on their own, and Meta's release of its personal AI agent Muse earlier last month. TechCrunch's coverage put Google's move in the same frame as consumer-facing agents like Meta's Muse and messaging-based operators.

The enterprise land-grab is already underway. On the same day as the launch, law firm Cooley announced it is teaming up with Google to co-develop a Gemini Enterprise agent for document redaction, joining what Law.com called a fast-growing slate of Big Law–Big Tech partnerships. And Tech Mahindra said on October 6 it will scale hands-on Gemini Enterprise knowledge across more than 12,500 associates.

All of this sits on top of Google's flagship model, Gemini 4 Argon, which the company unveiled earlier this month. The agent is the worker; Argon is one of the brains it can call on.

Our take: distribution beats demos

Here's the uncomfortable truth for the agent startups: Google has a distribution advantage that borders on unfair. A billion monthly users. The productivity suite half the world's companies already live in. The cloud contracts, the security certifications, the IT departments that already return Google's calls.

The agent wars won't be won by whoever has the cleverest demo. They'll be won by whoever can convince a chief information officer that an AI which works unsupervised inside the company's financial systems won't be a catastrophe. Google's entire pitch — businesses first, governance first, your rival's model if it's better — is aimed squarely at that CIO. It's a colossally boring pitch, and it's probably the winning one.

The open question is the one Google can't answer with a keynote: what happens the first time a "coworker agent" with its own email address does something its human colleagues would get fired for? The rogue-agent incidents we've been tracking suggest that day isn't far off. Google is betting that enterprise controls are the answer. Everyone placing that bet should read the incident reports first.

Sources: Reuters; TechCrunch; Unite.AI; Analytics Insight; Google Cloud Blog (Thomas Kurian keynote post); Law.com; Tech Mahindra announcement. Announcement date verified as October 8, 2026 across all outlets.

By Daniel Rao
Daniel Rao covers AI and enterprise tech for TechPulse USA.