Google's Gemini Agent Gets Its Own Email Address and Picks Claude When It Wants To
At Gemini at Work on Thursday, Google Cloud CEO Thomas Kurian announced "the Gemini agent," a single agent for work that answers questions, does knowledge work, makes media and writes and runs code from one prompt box. The line Google wants you to remember is "you give it objectives, not instructions." You delegate an outcome, close the laptop, and the work keeps running in the cloud for hours or days under one set of memories, no matter which device or channel you come back on. Businesses get it first; consumers later, because, as Sundar Pichai put it on stage, the hard problems are security, scale and performance. Gemini has over a billion monthly users and nearly 90% of the Fortune 100 use Gemini Enterprise, so the distribution question is already answered.
The architecture notes are the interesting part. The agent spins up temporary sub-agents, each with its own identity, for parallel and sequential steps. It can also run as a "coworker agent," a persistent role with its own @agents.company.com email, its own storage, and access only to the context a team gives it. It writes an audit trail attributed to itself, not to the person who tagged it. You reach it by tagging, emailing, sharing a doc, or adding it to a group chat, from iOS, Android, Windows, Mac, the command line, Workspace, Microsoft 365, ServiceNow or Slack. It connects to Confluence, Jira, Git, Salesforce, BigQuery, Databricks, Snowflake, Postgres and any MCP server inside or outside the network, and there is an enterprise registry for tools and another for skills, which Google defines as modular prompts for multi-step tasks, including skills the agent writes for itself.
Memory is four kinds: session, semantic, procedural and episodic. The agent "onboards itself the way a new hire would." And the model is a separate choice from the agent: it routes each job across the Gemini family and Anthropic's Claude models today, open and other private models later. Google's own argument for this is that "the leading model changes every few months," so the agent layer has to outlive any one model. PayPal routes 10 million multi-model requests a week; Shopify blends frontier models across millions of merchants. Pricing comes with multi-model orchestration, smart routing and real-time spend caps.
What Google just did is name the agent as the product and demote the model to a line item, which is the same move Anthropic made with Claude Code and OpenAI made with GPT-6's Intelligent UI last week. Three weeks ago a Wikimedia report asked for agents to be identifiable. Google's answer is an agent with an email address and an audit trail of its own. That is the identity layer the authority-gap thread has been waiting for, delivered as an enterprise feature. The open question is what happens when the coworker agent with its own inbox gets phished.
Links: cloud.google.com/blog/products/ai-machine-learning/welcome-to-gemini-at-work-2026, techcrunch.com/2026/10/08/google-brings-agentic-ai-to-gemini-starting-with-businesses/
← Back to all articles
The architecture notes are the interesting part. The agent spins up temporary sub-agents, each with its own identity, for parallel and sequential steps. It can also run as a "coworker agent," a persistent role with its own @agents.company.com email, its own storage, and access only to the context a team gives it. It writes an audit trail attributed to itself, not to the person who tagged it. You reach it by tagging, emailing, sharing a doc, or adding it to a group chat, from iOS, Android, Windows, Mac, the command line, Workspace, Microsoft 365, ServiceNow or Slack. It connects to Confluence, Jira, Git, Salesforce, BigQuery, Databricks, Snowflake, Postgres and any MCP server inside or outside the network, and there is an enterprise registry for tools and another for skills, which Google defines as modular prompts for multi-step tasks, including skills the agent writes for itself.
Memory is four kinds: session, semantic, procedural and episodic. The agent "onboards itself the way a new hire would." And the model is a separate choice from the agent: it routes each job across the Gemini family and Anthropic's Claude models today, open and other private models later. Google's own argument for this is that "the leading model changes every few months," so the agent layer has to outlive any one model. PayPal routes 10 million multi-model requests a week; Shopify blends frontier models across millions of merchants. Pricing comes with multi-model orchestration, smart routing and real-time spend caps.
What Google just did is name the agent as the product and demote the model to a line item, which is the same move Anthropic made with Claude Code and OpenAI made with GPT-6's Intelligent UI last week. Three weeks ago a Wikimedia report asked for agents to be identifiable. Google's answer is an agent with an email address and an audit trail of its own. That is the identity layer the authority-gap thread has been waiting for, delivered as an enterprise feature. The open question is what happens when the coworker agent with its own inbox gets phished.
Links: cloud.google.com/blog/products/ai-machine-learning/welcome-to-gemini-at-work-2026, techcrunch.com/2026/10/08/google-brings-agentic-ai-to-gemini-starting-with-businesses/
Comments