Google Cloud Launches Gemini Agent, Runs Claude Models

Google Cloud introduced Gemini agent on October 8, 2026 — a single agent that takes an objective instead of a step-by-step prompt and then does the work itself. The pitch is delegation rather than conversation: it writes code, runs tasks and carries one shared memory wherever you invoke it.

One agent, every channel

SiliconAngle reports access through the command line, Google Workspace, Microsoft 365, Slack and other third-party apps. The architectural claim that matters is persistence: a single set of memories that survives the jump between channels, so a job started in chat can be picked up on a desktop.

Google's own Product Hunt listing frames it as an agent that keeps running in the cloud for hours or days, and markets it with a line few vendors would write about a rival: the work agent that runs on both Gemini and Claude.

“You give it objectives, not instructions. You delegate an outcome and come back to finished work.” — Thomas Kurian, CEO, Google Cloud

Why Claude inside a Google product is the real story

Three tiers are named: Gemini Flash for short, simple jobs, the flagship Argon for long-horizon work, and Anthropic's Claude models, which SiliconAngle says are available today. Other leading private and open models are promised later.

No context window, no benchmark and no price is attached to Argon in the reporting. Multi-vendor routing is the concrete commitment here, and it cuts against the assumption that a hyperscaler's agent layer stays locked to its own weights.

Sub-agents with mailboxes of their own

Gemini agent can spawn sub-agents. Each gets its own identity, an address on the @agents.company.com pattern and its own persistent storage, with access limited to the context a user or teammate hands it.

Admin controls, audit logging and data residency are not described in the sources we could read. For regulated buyers, that is the gap to close before a pilot.

The billing problem, and Google's answer

An agent that runs for days is an open-ended invoice. Google's countermeasures are hard spend limits enforced in the Cloud Billing Console, smart routing that selects the cheapest adequate model per task, and per-project tracking so finance can allocate cost by department.

SiliconAngle also cites a roughly 98 percent fall in token prices since 2024, sourced to a linked third-party report rather than to Google. Read it as market context, not as a forecast for your own bill.

What the sources do not say

Availability status (general availability or preview), per-seat pricing and named launch customers are all absent from the material we could read. A third requested source, MarkTechPost, returned HTTP 403 on fetch and is therefore not cited below.