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

The most interesting part of Google's new Gemini agent sits underneath it. For each job, the agent picks a model, and the options already include Anthropic's Claude alongside Google's own Gemini family. Thomas Kurian, CEO of Google Cloud, announced the agent at Gemini at Work 2026 as a single, universal agent for work inside Gemini Enterprise, according to Google Cloud's write-up of the keynote.

At a glance

• Gemini Enterprise gets one agent and one API that answer questions, handle knowledge work, create images and media, and write and run code, working inline in Gmail, Drive, Docs, Sheets and the rest of Workspace.
• The agent spawns temporary sub-agents for multi-step jobs, keeps four kinds of memory, and can run as a coworker agent with its own @agents.company.com email address and persistent storage.
• Google pitches the agent on cost, citing Smart Routing, multi-model orchestration and real-time spend caps, yet the announcement names no prices for the agent or for those controls.

If you have not followed the renames: Gemini Enterprise is the successor to Google Agentspace. Google Cloud consolidated it under the new name in October 2025, with Agentspace technology at its core. According to Devoteam, Agentspace was Google's first platform for creating and managing agents, and Gemini Enterprise already offered a marketplace plus Google-made agents such as Deep Research, Data Insights and Code Assist.

Google says nearly 90% of the Fortune 100 already use Gemini Enterprise

Kurian opened with scale. In the last year, nearly 500 Google Cloud customers each processed more than one trillion tokens. Nearly 80% of all Google Cloud customers use its AI products, and nearly 90% of the Fortune 100 use Gemini Enterprise. Google's own framing is that work now starts in the prompt window.

The agent is built for that window. You give it objectives rather than instructions, and it plans the work, picks skills and tools, connects to company systems and returns something finished. It can act as your personal assistant, as a team member such as a project manager, or on behalf of a role, like a finance analyst.

The same keynote added plain-language data and analytics skills and industry packs of tools, skills and connectors for financial services and legal teams. Governance covers identity and policy management, permission controls, secure sandboxing and network gateways.

The model underneath can be Gemini or Anthropic's Claude

Google separates the agent from the model. Gemini is the agent, and the model is chosen per job. Today it orchestrates across the Gemini family and Claude models from Anthropic. Google says other leading private and open models will follow.

Google's stated reasoning is that the largest model is not always the best one. Matching the model to the task raises accuracy on hard jobs and cuts cost on simple ones. The leading model also changes every few months, Google says, so keeping the choice open means your context, skills and data stay put when it does.

Sportswear brand On tested the dynamic selection to get products to market faster. Google also points to Shopify, which blends frontier models for millions of merchants, and PayPal, which routes 10 million multi-model requests every week.

Coworker agents get their own @agents.company.com email addresses

The agent runs in the cloud, so it keeps one set of memories, one context and one personalization graph across web, iOS, Android, Windows, Mac, the command line, Workspace, Microsoft 365 and Slack. It can also run headless inside third-party apps, with no interface of its own. You can assign it work, schedule tasks or have it react to events, and jobs lasting hours or days keep running after you close the laptop.

For multi-step work, Gemini creates temporary, job-specific sub-agents, each with its own identity, and coordinates parallel and sequential steps. A coworker agent is the long-lived version: a defined role across days and sessions, its own @agents.company.com email and persistent storage, and access only to the context your team provides.

Inside Workspace, the same memory, skills and controls follow it into Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar.

Gemini keeps four kinds of memory and learns from shared registries

Three pieces do the work. Tools connect the agent to Confluence, Teams, Slack, Git, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Postgres, Snowflake and desktop files. It can also reach any Model Context Protocol (MCP) server, MCP being the open standard for plugging tools into AI agents. Skills are reusable instructions stored as modular prompts, and teams publish them, along with tools, to company-wide registries.

Memory comes in four kinds: session memory for the current task, even one running for days; semantic memory, a knowledge base built from documents, people and other agents; procedural memory for how a job gets done, including skills it writes itself; and episodic memory of everything it has done. Google compares it to a new hire who learns you, your tools and your team before starting.

Bradesco says document review fell from one hour to five minutes

The hardest numbers come from customers. Bradesco's finance team cut document review from 1 hour to 5 minutes, reduced risk inconsistencies by 60% and achieved over 10% in financial efficiencies. BNP Paribas is putting Gemini Enterprise into LLM@CIB, an internal assistant for more than 65,000 employees, for work like credit memos.

SOMPO built over 10,000 custom agents across its 34,000 employees, and its engineers cut new model development from one week to a day. Orange Spain runs more than 1,000 custom agents. At Wesfarmers, a Bunnings agent saved staff half a million hours, while shopping agents lifted conversion up to three times at Kmart and Officeworks.

Ulta Beauty's shopping assistant covers 30,000 products and tripled digital sales conversion. Santee Cooper, South Carolina's state-owned utility, expects financial modeling up to 75% faster across a $10 billion grid expansion budget. Merck is exploring molecular research.

Almost every number here comes from Google or its customers without baselines or methods, Santee Cooper's 75% is an expectation, and the keynote does not say which Claude models sit in the router. Oddly for an agent pitched on Smart Routing and real-time spend caps, it names no prices. The only figures available come from before this launch: Autenticare, a Google partner, reported US$21 (Business) or US$30 (Standard/Plus) per user per month after the Agentspace consolidation.

Which models join Claude next

Google says the router will add other leading private and open models in the future but has not named them or given a date. It has not published a price for the agent either. One related piece is already in Preview. Since October 2, 2026, federated query mode has let Gemini Enterprise query data where it sits, over MCP and with each user's own credentials, and nothing is copied into a separate store first.