Announced 8 Oct 2026 · Sources checked
What did Google announce at Gemini at Work?
The Google Cloud blog post “Welcome to Gemini at Work 2026: Introducing the Gemini agent,” dated 8 October 2026 and signed Thomas Kurian, CEO, Google Cloud, says it is adapted from his keynote. The editor’s note lists one agent for questions, knowledge work, images and media, and code; inline use inside Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar; data and analytics skills; financial-services and legal specializations; identity, sandboxing and network gateways; and cost controls through Smart Routing and real-time spend caps.
blog.google published a shorter English note the same day, “Google Cloud introduces the Gemini agent” (RSS 12:05 UTC), and points readers to the Cloud blog (RSS 12:00 UTC).
Kurian’s scale claims are Google’s: nearly 500 Google Cloud customers each processed more than one trillion tokens in the last year; nearly 80% of Google Cloud customers use Google’s AI products; nearly 90% of the Fortune 100 use Gemini Enterprise. Those figures are not independently audited here.
How does Google say the Gemini agent is built?
The keynote post frames Gemini as a single agent and a single API. “You give it objectives, not instructions.” It can be a personal assistant or a team member — a project manager for a group, or an analyst for a finance role. For the difference between an agent and a scripted workflow, see our explainer on AI agents.
Six architectural bullets sit under that claim. A unified agent answers in chat, works autonomously on assigned objectives, and generates code from one interface. Access is listed as web, iOS, Android, Windows, Mac, the command line, Workspace, Microsoft 365, Slack, and a headless mode inside third-party applications. Persistent execution runs in the cloud so memory and context follow the person across devices; jobs that take hours or days keep running after a laptop closes.
Multi-agent orchestration can spin up temporary sub-agents, each with its own identity, for parallel or sequential work. A coworker agent is described as a persistent teammate role, with its own @agents.company.com email, its own storage, and access only to context a person or team provides. Model choice treats the agent and the underlying model as separate: the post says Gemini family models and Claude models from Anthropic run today, with other private and open models later. The sportswear brand On, plus Shopify and PayPal, appear as early multi-model testers; PayPal is said to route 10 million multi-model requests a week. Those are vendor customer lines.
Named tools include Workspace, Microsoft 365, Slack, Git, Jira, Salesforce, ServiceNow, BigQuery and other data stores, desktop files, any MCP server, and an enterprise tools registry. Skills are reusable prompts in a global, company, or personal library. Memory is split into session, semantic, procedural and episodic stores. If you need the protocol the MCP sentence assumes, see Model Context Protocol.
Where does it run inside Workspace and data systems?
Inside Workspace the same memory, skills and controls are supposed to follow the agent into an email thread, a document and a Chat space. Three modes are listed: personal assistance that already knows calendars and related documents; proactive delegation, including a one-click hand-off when Workspace Intelligence flags a manager’s request for a slide deck; and a coworker agent that gets its own Workspace account and can be @mentioned. That is a different shape from Anthropic’s Claude sidebar beta for Docs, Sheets and Slides, which edits the open file rather than adding a directory identity.
Data skills split technical and business users. Engineers are told they can describe an outcome and get PySpark, notebooks, training and pipeline repairs. Business users are told they can ask for operational reports that Gemini constructs against BigQuery and a Knowledge Catalog, then rerun without extra token cost. Grounding is supposed to come from the Knowledge Catalog, Smart Storage on unstructured objects, and a “borderless Lakehouse” that queries Amazon S3 and Azure Data Lake without variable egress fees. Bloomberg Media is quoted as lifting SQL accuracy 63% during initial development; that is a customer quotation on Google’s page.
Industry packs are “now in preview for Financial Services and Legal, and coming soon to Government, Healthcare, and Retail.” The financial-services pack cites FactSet, LSEG, S&P Global and SEC filings, more than 50 foundational skills, and use by CME Group and Deutsche Bank. The legal pack cites NetDocuments, iManage, Harvey, Onit and a Cooley redaction agent. Those names and “preview” are Google’s.
How do you get it, and what is actually available?
The Cloud blog does not publish a price list or a general-availability date for the Gemini agent. The same-day Gemini for Business product page, titled “The Universal Work Agent,” says you can contact sales or try Gemini Enterprise Plus for 30 days without sales assistance. It also says: “Not every feature shown on this page is available right now.” Early-access items include multi-step background tasks, Gemini on mobile and desktop, and third-party model choice.
That last clause sits next to a contradiction you should keep. The keynote post says the agent already orchestrates Gemini family models and Claude. The product-page FAQ answers “Which models can I use?” with “Google Gemini models today, with additional models coming soon.” We are reporting both sentences. We did not open a console to see which router a trial tenant actually gets.
Cost controls in the keynote are Smart Routing, multi-model orchestration, and a hard project spend cap in Cloud Billing that pauses the project’s agent when token and sandbox costs hit the limit. Governance answers are identity, authorization through OAuth-style mapping, an audit trail attributed to the agent, an Agent Sandbox, and Agent Gateway as “an AI network firewall.” TPU 8i is said to deliver 80% better price-performance than the prior generation. Argon, Flash, Omni and Gemma are listed as the model lineup under the agent. That Argon name is the same family we covered when Gemini 4 Argon was announced; this post does not add a new parameter count.
What should readers not assume?
This is not a new foundation-model launch on the order of Gemini 4 Argon. It is a workplace-agent packaging of Gemini Enterprise, with a long customer appendix we have not verified. The product page already marks background jobs, mobile and desktop clients, and third-party model choice as early access. We did not sit the keynote, start the trial, or watch a coworker agent send mail from @agents.company.com.
Do not treat the customer hour-counts, conversion lifts or “10x faster” lines as measured results. Do not assume Claude routing is on in every tenant, or that industry packs beyond financial services and legal have shipped. Same-day Ireland, SMB and SUNY Cloud posts apply the same agent announcement; they are not separate product launches.
Common questions
Can the Gemini agent use Claude today?
The 8 October keynote post says yes: Gemini family models and Claude models are in the routing mix now, with other models later. The Gemini for Business FAQ says Google Gemini models today, with additional models coming soon, and it lists third-party model choice as early access. We did not resolve that in a live tenant.
Is this a new Gemini model?
No. The keynote describes an agent that sits on top of Argon, Flash, Omni, Gemma and, in some copy, Claude. It is a product and routing announcement, not a new parameter count.
How do you try it without a sales call?
The product page says Gemini Enterprise Plus has a 30-day trial without sales assistance. Several features on that page remain early access. We did not start a trial.
What to remember
Treat Gemini at Work as a first-party architecture announcement for one workplace agent, with a trial path and early-access flags on the product page. Keep Kurian’s customer numbers and the Claude-routing sentence in the vendor column until you can see them in your own tenant.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards
