DEV Community

Cover image for Google Cloud's Gemini agent can route work to Claude
techaiwire
techaiwire

Posted on Originally published at techaiwire.com

Google Cloud's Gemini agent can route work to Claude

#google #aiagents #anthropic #mcp

Google Cloud unveiled the Gemini agent on October 8, 2026, a single work agent that runs each job on whichever AI model fits best, including Anthropic's Claude. It is in private preview for business customers, according to 9to5Google. Google Cloud CEO Thomas Kurian presented it at the Gemini at Work 2026 event. Letting a rival's model run inside Google's own agent is a notable shift for enterprise AI.

What the agent is

Google calls it a "universal agent for work." It answers questions, does knowledge work, creates images and media, and writes and runs code, from one prompt box. The pitch is that you "give it objectives, not instructions," a line Yellow attributes to Kurian.

The agent splits long jobs among temporary sub-agents. These helper agents can work in parallel or in sequence. Google's blog says the work can run "for hours or days" and "keeps running after you close your laptop."

It works from the Gemini Enterprise app, inside Google Workspace, and through other channels. Reports list the command line, Slack and Microsoft 365. Inside Workspace, you can call it with @gemini in Gmail, Docs, Sheets, Slides and Chat.

Which models it uses

The agent picks a model for each job. It supports Google's Gemini models and Anthropic's Claude models today, with other private and open models planned, according to Google's blog. Google says its routing tool "automatically triages enterprise workloads so each one runs on the model that delivers maximum performance at the lowest possible cost."

Only Yahoo Tech calls this "Smart Routing." The other reports describe the same idea without that label.

Memory, identity and cost controls

Reports describe several controls that matter to businesses:

  • Four kinds of memory. Session memory covers the current task. Semantic memory holds knowledge. Procedural memory holds how to do things, including skills the agent writes for itself. Episodic memory holds past work.
  • Coworker mode. An agent can act as a team member with its own identity. Google's blog says it gets its own @agents.company.com email and persistent storage, and it sees only what team members share.
  • Headless use. The agent "operates as a headless agent, meaning it does not need a dedicated user interface," so other software can call it through an API.
  • Spending limits. Admins can set a hard limit on a project's AI spend. The agent pauses when it hits the limit and resumes with one click, Google's blog says.

What it connects to

Google's blog says the agent can read directly from Salesforce Data 360, SAP, ServiceNow and Workday without copying the data. It also lists Confluence, Jira, BigQuery, Databricks, Postgres, Snowflake and any MCP server. MCP is the Model Context Protocol, an open standard that lets an AI tool use other apps.

Industry versions are also planned. The blog says financial services and legal versions are in preview. Government, healthcare and retail versions are "coming soon." The financial services version has more than 50 foundational skills and uses data from FactSet and LSEG, according to Yahoo Tech.

Availability and numbers

The agent is in private preview. Google says wider availability is coming "soon" for Workspace customers on select Business and Enterprise plans. None of the sources gives a price or a general release date.

Kurian said nearly 80% of Google Cloud customers use its AI products, and nearly 90% of the Fortune 100 use Gemini Enterprise, per Google's blog. The 80% figure covers Google's AI products in general, not this new agent, as Yellow notes.

The same week, Anthropic opened Claude for Google Workspace in public beta, so Claude now appears in Google's tools from two directions.

What this means for developers

For teams building agents, the announcement shows where enterprise tools are heading.

  • Plan for several models. Google is routing between Gemini and Claude. Design your agent code so the model is a setting you can change, not a hard dependency.
  • Use spending caps as a safety net. A long-running agent can burn money for days. Set project limits before you let one run overnight.
  • Give agents their own identity. Coworker mode follows a sound pattern: one agent, one account, only the data it is shown. Avoid sharing a human's login with an agent.
  • Check your data connections. An agent that reads SAP or Salesforce directly needs tight permissions. Review what each connector can read and write.
  • Wait for a real trial. Private preview means limited access and no pricing. Treat the announced skills as plans until you can test them on your own work.

The details still missing are price and a release date. Until Google publishes them, budgeting for the Gemini agent is guesswork.

This article was first published on Tech AI Wire.

Also available in

Deutsch · 日本語 · Français · Español · Português

Related on Tech AI Wire

Sources

Top comments (0)