Google Wants Gemini to Be Your Next Coworker—Complete With Its Own Email Address

What it claims

Google Cloud CEO Thomas Kurian presented the Gemini Agent at the Gemini at Work event. It is a cloud-hosted agent that takes a goal, works on it for hours or days after the user closes their laptop, and returns a finished result. Managers can describe a role, such as "events coordinator", and get a coworker agent with its own Workspace account, email, calendar, Drive storage and directory listing. Decrypt says this is not new territory. It names comparable features from Grok, OpenAI and Hermes.

How it works

The page gives only the control surfaces, not the internals:

• Identity. Each agent gets a cryptographically attested identity. Its actions are logged to an audit trail under the agent, not under the manager who created it.
• Containment. Agents run in an Agent Sandbox. All of their traffic goes through Agent Gateway, which Google calls an "AI network firewall".
• Model routing. Gemini sends each job to either a Gemini model or an Anthropic Claude model, whichever Google judges the best fit for quality and cost. The page does not say whether customers can see or override that choice.
• Spend caps. Caps are set per project and enforced in real time. When a project hits its limit, the agent pauses and waits for a human to click resume. It does not fail or cut work short.

flowchart LR
  A[Agent task] --> B{Project spend cap hit?}
  B -- no --> C[Continue via Agent Gateway]
  B -- yes --> D[Agent paused]
  D -- human clicks resume --> C

Evidence

Google offers one scale figure: nearly 500 customers each processed more than one trillion tokens in the past year. The figure is about overall token volume on Google Cloud and is used to show why cost is a concern. It is not a measure of how the agent performs. The page has no benchmarks, task success rates or latency numbers.

Limits

• Not generally available. There is no published pricing and no GA date.
• In preview: financial services and legal versions.
• "Coming soon": government, healthcare and retail.
• Containment is asserted, not demonstrated. The article cites a July finding by Accomplish AI: Claude Cowork's local mode on Macs could escape its Linux VM and read files on the host. Google offers nothing showing that Agent Sandbox prevents this class of escape.
• Model routing is outside your control. Work may be processed by Anthropic models. Teams with data-residency or vendor constraints should confirm this before adopting.
• Pauses can stall long tasks. A task that hits its cap waits for a human. Plan who owns the resume click for multi-day work.
• The source is secondary. This is a news report on an event announcement. It has no documentation, API shape or setup instructions.