Google's Universal Gemini Agent Is Live in Early Access — a Direct Shot at Microsoft, OpenAI and Anthropic

MOUNTAIN VIEW, Calif.

Google's universal Gemini agent is live — in early access for enterprise customers since October 8 — and it is a direct shot at Microsoft, OpenAI and Anthropic. Three days after Thomas Kurian walked onto the stage at Hangar One and told the enterprise software world that "work now starts in the prompt window," the dust has settled, the analysts have filed their notes, and the picture is clear: Google didn't just ship a feature this week. It shipped an argument about how the AI agent race gets won.

The announcement, made October 8 at Google's Gemini at Work 2026 event with Google CEO Sundar Pichai sharing the stage, collapses Google's enterprise AI sprawl into one thing: the Gemini agent, which Kurian described as "a single, universal agent for work" that answers questions, handles knowledge work, creates images and media, and writes and runs code. "To meet this moment, today we are excited to announce the Gemini agent: your new single, universal agent for work," Kurian said in the company's launch post, as quoted by AI Weekly's coverage. "Work now starts in the prompt window."

The mechanics are the interesting part. You don't give it instructions; you give it objectives. The agent plans the work, spawns temporary sub-agents that run in parallel or in sequence, and keeps going after you close the laptop — progress shows up in a tasks inbox with the reasoning, delegated sub-tasks, loaded skills and written code all visible. That visibility matters: as Tech Insider's write-up of the launch noted, enterprise buyers wary of handing autonomous systems real responsibility need an auditable work trail, and Google is making the trail a feature rather than an afterthought.

Your coworker has an email address

The detail that traveled furthest this week is the coworker agent: a persistent, role-based agent that gets its own Google Workspace account — email address in the form @agents.company.com, calendar, Drive storage, and a seat in the company directory. Colleagues can add it to a Chat space, @mention it, or tag it in a Doc comment, where its edits appear under the agent's own name in version history. Every action is logged against the agent's cryptographically attested identity rather than a human user's, per the DEV Community's breakdown of the launch.

Underneath that identity sits four kinds of memory — session memory ("for the task in front of it, even when it runs for days," AI Weekly wrote in its coverage of the launch), plus semantic, procedural and episodic — so a job that takes a week doesn't lose context every time a chat window closes. It works inline across Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, and reaches outside Google's walls into Microsoft 365, Slack and Teams, from web, mobile, desktop and the command line. The tools registry covers Salesforce, ServiceNow, Jira, Snowflake, Databricks, BigQuery, Git and any MCP server, according to the Futurum Group's analysis of the announcement.

This is the part Microsoft should be studying closely. Redmond owns the distribution — Copilot sits inside the Office apps where work already happens — but Google just made the case that the agent itself should be the interface, reaching into Microsoft 365 rather than living inside it. That's a direct challenge to the home-field advantage.

The Claude gambit

Here's the move that surprised people: the Gemini agent doesn't just run on Google models. It routes each job across more than 200 models, including Google's own Gemini family — Argon for frontier reasoning, Flash for speed, Omni for generative media, Gemma for lightweight edge workloads — and Anthropic's Claude, with other private and open models planned. A Smart Routing tool picks which model handles each job, per AI Weekly's coverage of the launch.

Google's argument is explicit: the biggest model isn't always the best one for a given job. And the subtext is aimed at everyone: at OpenAI, whose enterprise pitch is built on its own models; at Anthropic, whose Claude is now inside a Google product (awkward, but lucrative); and at the whole idea of model lock-in. Google is betting that enterprises don't want to marry a model — they want an agent that shops around. Given that Anthropic just shipped Claude Sonnet 5.5 and OpenAI is pushing GPT-6 Luna, the timing of the agnostic pitch is not accidental.

The governance layer is where Google thinks it wins the RFP. Cryptographically attested agent identity. OAuth-propagated permissions, so the agent only sees what you share with it. Action-level audit trails attributed to the agent, not a human. An Agent Sandbox and an Agent Gateway that Google calls an AI network firewall. Real-time, per-project spend caps in the Cloud Billing Console that pause an agent when triggered. On the data side: a Knowledge Catalog for business definitions, Smart Storage for unstructured data enrichment, and a Borderless Lakehouse that queries Amazon S3 and Azure Data Lake with no variable egress fees — plus financial services and legal specializations in preview, per Futurum. This is a product designed by people who have sat through enterprise security reviews. It shows.

What "soon" actually means

The honest caveats, because there are several. Pricing hasn't been announced. There's no general-availability date — the agent is in private preview / early access, with Google saying 50+ companies are running it today, and 9to5Google reporting wider availability "soon" for Workspace customers on select Business and Enterprise plans. A company representative said there are no plans for per-seat licensing of coworker agents in Workspace, but the final metering model is undisclosed. Pichai said a consumer version is coming soon, which is the part that should interest everyone who doesn't have a CIO.

Wall Street noticed. Barchart's take this week framed the launch as a potential catalyst for Alphabet stock, which had spent nearly five months tightening into a textbook base — the question being whether a product unveiling is enough to light the fuse. The market's answer so far: intrigued, not convinced. Which is fair. Enterprise software is sold in quarters, not keynotes.

And the race is genuinely a race now. Microsoft has distribution. OpenAI has mindshare and the consumer on-ramp. Anthropic has the model that even Google will route your work to. Google's bet is that none of that matters as much as the boring stuff — identity, permissions, audit trails, spend caps — because the boring stuff is what lets a CIO say yes. It's the least glamorous theory of victory in the agent wars, and it might be the right one.

More from the newsroom: our launch-day coverage (Oct 8) · Anthropic's Claude Sonnet 5.5 launch · OpenAI's GPT-6 Luna rollout · Apple's full-disk-access AI agents and the security questions