Google Cloud Launches ‘Gemini’ Agent for Work, a Universal AI Coworker With Its Own Email Address

Abdul Mannan 2 hours ago • 9 min read

A glowing holographic AI agent figure made of blue light working alongside office professionals at desks with computer monitors and floating holographic charts in a modern office

Google Cloud has unveiled what it calls a universal agent for work, a single artificial intelligence system named Gemini that the company says can answer questions, handle knowledge work, create images and media, and write and run computer code, all from one prompt box. The announcement, made at the company’s Gemini at Work 2026 event on Thursday, marks Google’s most aggressive move yet to collapse the growing sprawl of workplace AI tools into one autonomous assistant, according to Google Cloud.

With the new agent, work “now starts in the prompt window,” Google said in its announcement on The Keyword, the company’s official blog. Rather than toggling between a chatbot for questions, a separate tool for document drafting and another for code, an employee can type one instruction and have Gemini plan the work, choose the appropriate software tools, connect to the company’s internal business systems and return finished work inside the documents, inboxes and developer environments people already use, according to the company.

The launch places Google Cloud squarely in what is becoming the AI industry’s most contested race: not who builds the smartest model, but who sells the first agent businesses actually trust to do work on their own. Alphabet shares were up marginally in early trading on Thursday, according to Reuters.

“Gemini has all of an organisation’s business context and can be used for everything from knowledge work to answering questions, and content creation to coding, all from a single prompt box,” Google said in its announcement. The company emphasised that the agent plans the work itself, uses skills and tools rather than just generating text, and delivers finished outputs rather than drafts that require further assembly.

A digital coworker with a desk — and an email address

The most striking detail of Thursday’s announcement is how far Google is willing to let the agent impersonate a colleague. Users can create what Google calls “coworker agents” that act as persistent members of a team, each with its own email address, calendar, storage drive and presence in the company directory, according to Reuters’ account of the launch.

An example given at the event: a manager could create an “Event Planner Agent” that receives its own Workspace account, email address, calendar, Drive storage and company-directory entry. Employees would then assign work to it much as they would to a colleague, including by mentioning it in a chat message or requesting document edits, according to Google Cloud chief executive Thomas Kurian’s keynote post on the Google Cloud Blog.

The coworker configuration is designed for continuity. Because execution happens in the cloud, an employee can start an assignment on one device, leave it running and resume from another without reconstructing the context, according to the company. Google says the agent can manage workflows lasting multiple days, coordinate parallel assignments, respond to scheduled tasks or events, and even assemble temporary teams of specialised sub-agents for complex work, distributing different responsibilities among them.

These agents operate through a common interface and application programming interface, with access extending across the web, iOS and Android devices, Windows and Mac computers, command-line tools, Google Workspace, Microsoft 365 and Slack, according to reporting on the announcement. The agent works inline directly inside Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, carrying the same memory, skills and controls everywhere, according to the Google Cloud Blog.

The multi-model bet: Gemini picks the winner, even when it is Claude

Perhaps the most consequential technical detail is that Gemini does not pledge allegiance to Google’s own models. The agent picks the best model for each task, and at launch it runs on Google’s Gemini models and Anthropic’s Claude models, with other models to be added later, according to Reuters.

Sportswear brand On was cited at the event as an early tester of the dynamic model-selection capability, while payment giant PayPal now routes 10 million multi-model requests every week and Shopify blends frontier models for millions of merchants, according to reporting on Kurian’s keynote. The architecture treats models as interchangeable engines behind one interface, a design Google says enables “multi-model orchestration” and “Smart Routing” to control spending.

That plumbing extends to cost. The announcement promises built-in cost controls — multi-model orchestration, real-time spend caps and routing to cheaper models where possible — alongside the security, administration and governance controls that enterprise customers demand, according to Google. The governance layer includes identity and policy management, authorisation and permission controls, secure sandboxing and network gateways, the Google Cloud Blog said.

Google also announced industry-specific versions of the agent, with editions built for financial services and legal work now available in preview, and versions for government, healthcare and retail coming soon, according to Reuters. A separate set of new data and analytics skills lets both technical teams and ordinary business users ask plain-language questions and receive actionable operational insights within minutes, according to the Google Cloud Blog.

Big numbers, and bigger customers

Kurian’s keynote leaned heavily on adoption statistics that Google clearly intends as evidence the enterprise AI market has crossed from experimentation to infrastructure. Nearly 500 Google Cloud customers each processed more than one trillion tokens over the preceding year, according to the keynote as reported by Unite.AI. Nearly 80 per cent of all Google Cloud customers are now using its AI products, and nearly 90 per cent of the Fortune 100 use Gemini Enterprise, the company’s workplace AI platform, the same report said.

Gemini Enterprise is Google’s platform for bringing AI to every employee across every workflow, announced by Google chief executive Sundar Pichai at a Google Cloud event and positioned as the new front door for AI in the workplace. Industry-specific editions for financial services and legal launched in preview in August, each adding purpose-built skills, data connectors and third-party agent partners, according to reporting on the earlier launch.

The keynote’s named proof points spanned retail, payments, banking and telecoms. Brazilian bank Bradesco cut document review time from one hour to five minutes, while telecommunications operator Orange Spain has deployed more than 1,000 custom Gemini Enterprise agents, according to the keynote reporting. Shopify and PayPal were cited as blending multiple frontier models at merchant scale.

The race is now about autonomy, not answers

Thursday’s launch lands in the middle of an industry-wide pivot from chatbots to agents that act. OpenAI launched always-on agents in September that chase user goals across apps on their own, while Meta released its personal AI agent, called Muse, in September, which can shop, book travel, send emails and make payments for users, according to Reuters.

The pattern across all three announcements is unmistakable: the largest AI companies have concluded that the future of workplace AI is not a smarter chat window but a persistent digital worker that plans, uses tools and finishes jobs. Google’s version distinguishes itself by aiming that worker directly at the enterprise stack — company data, governance policies, existing identity systems — rather than at consumers.

Analysis: Why It Matters

The consolidation story is the headline beneath the headline. For two years, the enterprise AI market has fragmented into dozens of copilots — one for email, one for code, one for spreadsheets, one for support tickets. Each was competent; collectively they recreated the very tool sprawl they promised to eliminate. Google’s answer is to invert the model: instead of bringing AI to each application, bring the worker to the AI. The prompt box as the new desktop is a bold interface bet, and it puts Google in direct collision with both Microsoft’s Copilot strategy and OpenAI’s own enterprise ambitions.

But the genuinely disruptive detail is the multi-model routing. A Google agent that can choose Anthropic’s Claude over Google’s own Gemini is a quiet admission that no single lab will win every task — and, more importantly, that customers will not believe a single-lab answer is honest. By letting the agent route to the best model per job, Google trades tribal pride for credibility, and it buys itself a strategic hedge: if a rival model leads in some category, Gemini still delivers the answer. It also puts quiet pressure on the entire industry to stop grading agents by whose logo sits on the model and start grading them by outcomes per pound, dollar and rupee.

Then there is the email address. Giving a coworker agent its own inbox, calendar and directory entry sounds like a marketing flourish, but it is really a governance architecture. In enterprise IT, identity is the unit of accountability: permissions, audit trails, access revocation and compliance all hang off who someone is. An agent with an identity can be onboarded, monitored and fired like an employee. An agent without one is shadow IT. Google’s framing of identity, authorisation and sandboxing as headline features — not footnotes — suggests the company has learned that the enterprise objection to autonomous agents was never really about intelligence; it was about control.

The customer statistics deserve a sceptical eye even so. Processing trillions of tokens is a measure of throughput, not of value delivered, and “using AI products” can mean anything from a single pilot to full deployment. Google’s numbers are directionally impressive but they are marketing numbers, self-reported from a keynote. The more interesting signal is the vertical-specific versions: financial services and legal first, government and healthcare next. Those are the industries where errors have regulators attached — precisely the customers who will not tolerate an agent that hallucinates in a loan file. If banks trust it, the trust question is being answered where it is hardest.

What to watch next is threefold. First, pricing and general availability: the announcement is light on when ordinary customers get the full coworker configuration, and enterprise buyers will want the spend-cap mechanics spelled out before they hand an agent the keys. Second, the rivals’ response: OpenAI’s dots and Meta’s Muse now face a Google offering with deeper enterprise plumbing, and Microsoft — absent from Thursday’s headlines but present in every Office suite — cannot stay quiet for long. Third, the labour question that hovers over every coworker-agent demo: an agent with an email address is one abstraction layer away from a headcount line. Google frames it as augmentation; workforces will hear replacement. How companies disclose and manage that will define the politics of AI in 2027.