Skip to content

Google

Google’s Gemini agent can run jobs on Anthropic’s Claude

7 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: threads.com

Google Cloud CEO Thomas Kurian used his October 8 keynote at Gemini at Work 2026 to introduce the Gemini agent, a single agent for Gemini Enterprise. Under the Gemini name, the agent and the model are two separate things. Each job goes to whichever model fits it, and today that means Google’s Gemini family or Anthropic’s Claude, according to the Google Cloud post adapted from the keynote.

One prompt box and one API cover answering questions, knowledge work, image and media generation, and writing and running code. Kurian’s pitch fits in six words: “You give it objectives, not instructions.” You hand off an outcome and come back later to finished work, delivered in the documents, inbox and developer environments you already use.

One agent, several ways in

The same agent handles chat, autonomous work toward an assigned objective, and code generation from one interface. Work can also be scheduled, or set to start when an event fires. It runs on the web, on iOS and Android, on Windows and Mac desktops, and through channels that include the command line, Google Workspace, Microsoft 365 and Slack. It can also run headless, without a user interface of its own, so other applications can embed it.

The agent runs in Google’s cloud, not on the device you’re typing on. It keeps one set of memories, one context and what Google calls one personalization graph, whichever device or channel you use. Google says jobs that take hours or days keep going after you close your laptop, and you never have to re-brief the agent on another machine.

In Workspace it works inline in Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, with the same memory, skills and controls it has everywhere else. Google also announced data and analytics skills that let both technical teams and business users get operational answers from plain-language questions, along with tools, skills, connectors and domain knowledge for financial services and legal teams.

Sub-agents and coworkers

The agent comes in two multi-agent forms. For a multi-step task it can spin up a roster of sub-agents. These are temporary and built for one job, but each has its own identity. It coordinates them through steps that run in parallel or in sequence, and a workflow can last hours or days.

The second form is the coworker agent, which Google describes as closer to a team member. It has a fixed role and stays present across days, sessions and changing responsibilities. Coworker agents get their own @agents.company.com email addresses and their own persistent storage, and they see only the context that you or your teammates give them. Gemini can act as one of these, delegating and coordinating tasks the way a project manager would for a team, or covering a specific role such as an analyst in the finance department.

Google says it governs these agents through identity and policy management, authorization and permission controls, secure sandboxing and network gateways.

What does it remember?

The keynote names three things that let the agent work like an insider: tools that connect it to company systems, skills that describe how work gets done, and context that lets it remember.

On tools, the list includes Confluence, Microsoft Office, Teams, Slack and Workspace; Git and Jira; Salesforce and ServiceNow; BigQuery, Databricks, Postgres and Snowflake; and files on your own desktop. The agent can also reach any Model Context Protocol (MCP) server, the open standard for plugging tools and data sources into an AI assistant, whether that server sits inside or outside the company network. For comparison, custom MCP server connections reached the Business edition of Gemini Enterprise as a public preview in September. Teams can publish their own tools to an enterprise tools registry.

Skills are reusable instructions, knowledge or workflows stored as modular prompts that teach an agent a specific multi-step task. Gemini ships with a global library of them. Departments can publish custom skills to a shared company registry, and individuals can write personal ones. Google says the agent picks the skills and tools a task needs and learns from each run, which makes it more consistent and cuts token costs.

Memory is split four ways:

  • session memory for the task in front of it, even one that runs for days;
  • semantic memory, a structured knowledge base it builds as it reads documents, talks to people and works with other agents;
  • procedural memory for how a job gets done, including skills it writes for itself;
  • episodic memory of everything it has done before.

Teams can set up dedicated projects to narrow the context, skills and tools the agent uses for their work. In Google’s words, Gemini “onboards itself the way a new hire would,” learning the person, their tools and their team before it starts.

Why Claude sits underneath

Google describes the model layer as orchestration. The agent runs each job on the model that fits best, drawing on Gemini models and Claude models from Anthropic now, with “other leading private and open models in the future” promised. The keynote argues that matching the model to the work raises accuracy on hard jobs and lowers cost on simple ones. “The best model for the task is not always the largest one,” it says. It adds that the leading model changes every few months, and keeping the choice open means context, skills and data stay where they are when that happens.

Cost controls are part of the pitch too: multi-model orchestration, Smart Routing and real-time spend caps. Prices for the agent itself were not given. As evidence that multi-model setups already work at scale, Google points to customers. Sportswear brand On tested dynamic model selection to get products to market faster, Shopify blends frontier models for millions of merchants, and PayPal routes 10 million multi-model requests every week.

The scale Google is selling into

Kurian put numbers on the installed base. In the past 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. Google launched Gemini Enterprise on October 9, 2025 as the successor to Agentspace, combining Gemini models, the Agentspace framework and internal and external agents in one platform. The new agent arrives almost exactly a year later.

The customer figures in the keynote come from Gemini Enterprise deployments in general, not from the new agent:

  • Bradesco cut document review time from 1 hour to 5 minutes in fiscal, accounting and contract analysis, reduced risk inconsistencies by 60% and achieved more than 10% in financial efficiencies.
  • SOMPO built more than 10,000 custom agents across its 34,000 employees, and its engineers cut development time for new models from a week to a day.
  • Orange Spain deployed more than 1,000 custom agents across HR, IT, sales and customer service.
  • BNP Paribas is bringing Gemini Enterprise into LLM@CIB, its internal assistant for more than 65,000 employees, for work such as preparing credit memos.
  • Ulta Beauty’s shopping assistant covers 30,000 products, and Google says it tripled digital sales conversion rates.
  • At Wesfarmers, an internal Bunnings agent saved staff half a million hours of administrative work, and shopping agents raised conversion rates up to three times for Kmart and Officeworks.

Santee Cooper, South Carolina’s state-owned utility, expects financial scenario modeling to run up to 75% faster with a specialized model built on Gemini Enterprise. It is using it to manage market risk across a $10 billion grid expansion budget.

Later, on October 15, Google’s Gemini Enterprise release notes listed Claude Opus 5.5 and Claude Sonnet 5.5 as generally available in the Gemini Enterprise app. They run in the global location and the us and eu multi-regions and are billed pay-as-you-go. An administrator first turns a model on under “Enable third-party models” in the Feature Management tab, and then users can pick it in the model selector.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.