Gemini agent: Google Cloud wants one agent for all your work

One agent, one window, one API

Thomas Kurian's message fits in one sentence: work now starts in the prompt window. The Gemini agent is pitched as the single entry point for answering questions, producing documents, creating images, and writing and running code.

The promise is delegation: you set a goal, not a sequence of instructions, and come back to finished work. The agent can be triggered on demand, scheduled, or reactive to events.

It can act as a personal assistant, as a team member (a project manager, for instance), or on behalf of a specific role in the organization, like a financial analyst.

Architecture: persistence, sub-agents, and model choice

The agent runs in the cloud and keeps a single personalization graph regardless of device. It draws on four kinds of memory: the current session, a semantic knowledge base, procedural memory (including skills it writes for itself), and an episodic memory of everything it has already done.

For long tasks, it spins up a team of temporary sub-agents, each with its own identity, and coordinates steps in parallel or sequence. "Coworker agents," by contrast, are durable, have an address at @agents.company.com, and only see the context they're explicitly given.

The most strategic point: Google decouples the agent from the model. Gemini and Anthropic's Claude are orchestrated together today, on the grounds that "the best model for the task is not always the largest one" and that the leading model changes every few months.

The agent plugs into existing tools (Confluence, Teams, Slack, Jira, Salesforce, ServiceNow, BigQuery, Snowflake, Databricks…) and any MCP server, with enterprise registries to share tools and skills.

Inside Workspace, a coworker with an account

In Gmail, Docs, Sheets, Slides, Chat and Calendar, Gemini works inline. One example given: scheduling a meeting with "the usual team of regional managers" without providing a single name, the agent inferring participants from a Chat space and an old thread.

Delegation becomes proactive: when a manager asks for a slide-deck status update, Workspace Intelligence offers to hand the task to Gemini with one click, and sorts the inbox by importance rather than by date.

A coworker agent created from a simple role description gets email, calendar, Drive and a directory entry. It comments on documents under its own name and shows up in version history.

Data and vertical use cases

For data and ML engineers, Gemini generates PySpark, trains models, and fixes pipelines. For business users, it builds reporting queries on BigQuery that, once saved, run without any token cost.

Three building blocks ground the answers: the Knowledge Catalog (shared business definitions, which pushed SQL accuracy up 63% at Bloomberg Media), Smart Storage for the 90% of data that's unstructured, and a Borderless Lakehouse that queries S3 and Azure with no egress fees and federates Apache Iceberg tables.

The Finance and Legal versions are in preview. The former touts confidence scores, data lineage and verifiable citations, already used by CME Group and Deutsche Bank; the latter inherits ethical walls from iManage and NetDocuments, with Harvey, Onit and Cooley as partners.

Governance and cost, the two conditions for scale

Kurian boils governance down to four questions: who is the agent, what can it do, what has it done, what must it never touch. The answer: cryptographically attested identity and least privilege, permissions propagated via OAuth, and audit attributed to the agent rather than to a human.

Everything runs inside an Agent Sandbox, and all traffic passes through Agent Gateway, a network firewall for AI. A policy written once ("no access to Need to Know documents") applies to every agent.

On cost, Google notes that price per token has dropped 98% since 2024, but volumes have exploded. Hence multi-model routing, Smart Routing, and per-project spending caps that pause the agent, with the option to charge back by department.

Infrastructure, models, and customers

The TPU 8i is said to deliver 80% more price-performance than the previous generation. The model lineup is spelled out: Argon for frontier reasoning, Flash for speed, Omni for generative media, Gemma for edge use in open weights, which NASA's JPL runs on a satellite in orbit.

The customer list is long: BNP Paribas (65,000 employees), Bradesco (document review cut from one hour to five minutes), SOMPO (over 10,000 agents), DBS (chains of 70 to 80 agents for credit memos), and the U.S. CDAO, which put Gemini Enterprise in the hands of 3 million military personnel.

Accenture is creating a dedicated business group, and consulting partners have run more than a hundred thousand consultants through Gemini hackathons in a single month.