Google Launches the Gemini Agent for Work: One Agent, Claude Models Included

Google Cloud has launched the Gemini agent, a single cloud-hosted agent for work that answers questions, does knowledge work, generates media, and writes and runs code, all from one prompt box. The announcement came from CEO Thomas Kurian's keynote at Gemini at Work 2026 on October 8. It is in private preview for enterprise customers, and it can run Claude models from Anthropic underneath alongside Gemini. This is Google's clearest statement yet that the agent, not the model, is the product.

If you have followed Google Cloud Next 2026 and the Gemini Enterprise agent platform, this is the layer that sits on top of it. Below is what Google actually announced, what is still a claim, and what builders and IT teams should do with it.

TL;DR: the Gemini agent at a glance

What did Google actually announce?

Kurian framed the launch with usage numbers from Google's own post: in the last year nearly 500 Google Cloud customers each processed more than one trillion tokens, nearly 80% of Google Cloud customers use its AI products, and nearly 90% of the Fortune 100 use Gemini Enterprise. These are Google's figures and are not independently verified.

The announcement itself, in Google Cloud's keynote write-up, rests on a handful of design principles:

• Unified agent. The same agent answers chat questions, works toward objectives you assign, can be scheduled, and can respond to events.
• Omnipresent access. It is reachable from any device or channel and can run headless inside third-party apps.
• Persistent execution. Work that takes hours or days keeps running after you close your laptop.
• Multi-agent orchestration. It can create temporary sub-agents, each with its own identity, and coordinate parallel and sequential steps.
• Deep context. It keeps session, semantic, procedural, and episodic memory, and Google says it "onboards itself the way a new hire would."
• Model choice. The agent is separate from the model underneath it.

The Verge's report adds the important caveat: the agent is currently available only to enterprise customers in private preview. It also notes that this sits beside the consumer-facing Gemini Spark agent launched in May, which lives in a separate interface within the Gemini app.

Why the Claude detail matters

The most striking line is about models. Google says the agent runs "each job on the model that fits best," orchestrating across Gemini models and Claude models from Anthropic today. The stated logic: the best model is not always the largest, and "the leading model changes every few months," so keeping the model swappable means your context, skills, and data stay put.

That is a notable admission from a lab with its own frontier line. We covered the same idea at the developer level in putting Claude Fable 5.1 and Gemini 3.8 Flash on one coding team, and the cost logic echoes Gemini context caching and agent token costs. Google cites early use of the dynamic selection by the sportswear brand On, and points to Shopify blending frontier models and PayPal routing 10 million multi-model requests per week as proof that multi-model strategies already work at scale.

For a buyer, the practical read is that vendor lock-in to a single model family is explicitly not Google's pitch. For Anthropic, it is another distribution channel next to its own Claude Managed Agents and enterprise managed auth work.

Coworker agents: an agent with its own email

The feature most likely to get argued about is the coworker agent. You describe a role, such as an events coordinator, and Gemini creates an agent with:

• a dedicated identity, including an @agents.company.com email address
• its own Workspace account: calendar, Drive, and a presence in the company directory
• persistent storage
• access only to the context you or team members provide

Colleagues work with it like anyone else: add it to a Chat space or @mention it. Google's example is a marketing manager asking an events coordinator agent to draft a launch readiness document, which the agent posts back to the group. In a Doc comment thread it can suggest an edit and reply, "appearing under its own name in version history."

Google stresses that the agent acts under its own identity rather than yours, and that access follows the sharing and membership your team already uses. That is the right design question, because the failure mode of agents acting as you is well documented. See our coverage of AgentCorruption and hijacked Bedrock AgentCore agents for what happens when agent identity and permissions are loose.

Inline in Workspace: three modes

Inside Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar, Google describes three ways the agent works:

1. Personal assistance. It arrives knowing your calendar, team, and projects. Asked to set up a meeting with the usual regional event leads, it infers the people from the chat space and last thread, checks calendars, and starts an email thread, even with external participants.
2. Proactive delegation. Workspace Intelligence spots a delegable task, such as a manager asking for a project update as a slide deck, and offers a single click to hand it to Gemini.
3. A member of your team. The coworker agent described above.

Cross-app chains are the pitch: research market trends, build a financial model in Sheets, then produce a deck, without re-explaining the project at each step.

Tools, skills, and MCP

Google positions three building blocks. Tools connect to systems including Confluence, Microsoft Office and Teams, Slack, Git, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Postgres, Snowflake, and files on a desktop, plus any Model Context Protocol server inside or outside the company network, with an enterprise tools registry. Skills are reusable instruction bundles, a global library plus a shared company registry and personal skills. Context is the memory stack described earlier.

If those terms are new, start with our explainers on what MCP is and what agent skills are. The notable thing is that Google adopted the skills-registry and MCP vocabulary the open ecosystem already uses, instead of inventing a proprietary packaging format.

Data and analytics skills

The first domain skills target data work. For engineers, Gemini generates PySpark, provides notebooks, trains models, and troubleshoots pipelines. For business users, it builds operational reports via BigQuery and the Knowledge Catalog, saves the query, and lets teams rerun reports "without incurring token costs." That saved-query design is a sensible cost lever: pay tokens once to author the query, then run plain SQL.

Google also announced Smart Storage, which enriches unstructured files in place, and a Borderless Lakehouse that queries Amazon S3 and Azure Data Lake without variable egress fees and federates Iceberg tables across Databricks Unity, Snowflake Polaris, and AWS Glue. Bloomberg Media reported a 63% lift in SQL query accuracy during development after grounding its agents in the Knowledge Catalog, a customer figure from Google's post.

Industry specializations

Gemini for Financial Services is in preview, drawing on FactSet, LSEG, SEC filings, and proprietary data, with more than 50 foundational skills and outputs that show confidence scores, methodology, data lineage, and citations. CME Group and Deutsche Bank are named users. The legal pack inherits matter-level permissions and ethical walls from document platforms such as NetDocuments and iManage. Government, healthcare, and retail are listed as coming soon.

Anthropic is pushing the same vertical angle with Claude financial services agents and plugins, so expect these two to be compared head to head by bank and law-firm buyers.

Customer claims, read carefully

Google's post lists many outcomes. Treat them as vendor-reported:

Several of these are built on Gemini Enterprise broadly, not necessarily on the new agent, which has only just entered private preview. Do not read them as Gemini agent results.

What is claimed versus what is verified

We found no published pricing for the agent in the keynote post, and no general availability date. Both are the first things to ask a Google rep.

What this means for what you build or pay

• Procurement: the agent layer is being unbundled from the model. Ask whether skills, memory, and registries are portable if you change models, since that is Google's own argument for the design.
• Security: a coworker agent with its own mailbox is a new identity type. Plan for how it is provisioned, audited, and offboarded, and review the sandboxing, network gateway, and policy controls Google says it ships.
• Cost: spend caps and Smart Routing are promised, not proven. Pilot with a hard budget, and compare against the multi-model patterns we describe in our Gemini 3.8 Flash coverage.
• Naming: Google's product names keep shifting; our Google AI product names glossary helps keep Gemini, Gemini Enterprise, and the agent straight.

What to watch next

Three signals will tell you whether this is real. First, general availability and pricing. Second, whether independent testers confirm that routing to a Claude model actually lowers cost without hurting quality. Third, whether coworker-agent identity and permission models survive adversarial testing. We will update this post as Google publishes documentation and as hands-on reports appear.

Related reading

• Google Cloud Next 2026: TPU 8 and the Gemini Enterprise agent platform
• Claude Fable 5.1 and Gemini 3.8 Flash on one coding team
• Gemini context caching and agent token costs
• Claude Managed Agents: dreaming and multi-agent orchestration
• What is MCP? The Model Context Protocol guide
• What are agent skills? Complete guide
• Google AI product names, 2026 glossary

Details reflect Google Cloud's announcement and The Verge's report as of October 8, 2026; availability and features may change.