Google Cloud announced the Gemini agent at its Gemini at Work 2026 keynote, adapted from Thomas Kurian’s address. Gemini is positioned as a single, universal work agent: it answers questions, performs knowledge work, creates images and media, and writes and runs code from one prompt box. It works inline inside Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar, retaining the same memory, skills, and controls everywhere, and can also run on mobile, desktop, and third-party channels or as a headless agent. The intended interaction model is objective-based: users delegate an outcome and receive finished work rather than giving instructions. Inside Workspace, it can act as a personal assistant, proactively suggest delegatable tasks, or operate as a coworker agent with its own account, email, calendar, Drive, and directory presence that colleagues can @mention.

The architecture rests on six principles. A unified agent handles chat, autonomous task execution, and code generation from one interface. Omnipresent access covers web, iOS, Android, Windows, Mac, Workspace, Microsoft 365, Slack, and third-party apps. Persistent execution keeps a single memory, context, and personalization graph in the cloud, so long-running jobs continue after a laptop closes. Multi-agent orchestration lets Gemini create temporary sub-agents or persistent coworker agents with their own @agents.company.com emails, storage, and access boundaries. Deep context comes from tools, data, work history, and per-team projects. Model choice flexibility decouples the agent from the model underneath: Gemini can route across its own Gemini family and Anthropic’s Claude models today, with other private and open models planned, matching model size to task difficulty to improve accuracy on hard jobs and lower cost on simple ones. Early testers include On, and existing large-scale multi-model strategies include Shopify blending frontier models for millions of merchants and PayPal routing 10 million multi-model requests every week.

Capabilities are organized into tools, skills, and context. Tools connect Gemini to Confluence, Microsoft Office, Teams, Slack, Workspace, Git, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Postgres, Snowflake, desktop files, and any MCP server, with an enterprise tools registry for publishing tools. Skills are modular prompt-based workflows in a global library that teams and individuals can extend. Context uses four memory types: session, semantic, procedural, and episodic. For data and analytics, Gemini generates PySpark code, notebooks, model training, and pipeline fixes from plain-language outcomes; business users can create grounded operational reports from natural language, with saved queries avoiding token costs. Grounding is supported by Knowledge Catalog for business definitions, Smart Storage for enriching unstructured data in place, and a Borderless Lakehouse that queries Amazon S3, Azure Data Lake, Databricks, Snowflake, SAP, ServiceNow, and Workday without data movement.

Industry specializations are in preview for Financial Services and Legal. Gemini for Financial Services uses trusted data from FactSet, LSEG, SEC filings, and proprietary repositories, with more than 50 foundational skills, confidence scores, methodology disclosure, data lineage, and citations; CME Group and Deutsche Bank are named users. Gemini for Legal inherits matter-level permissions and ethical walls from NetDocuments and iManage, with Harvey, Onit, and Cooley building on it. Deloitte’s Agentic AI blueprint claims up to a 3x faster path to measurable business impact via operational heat maps.

Security and cost governance are treated as prerequisites for enterprise deployment. Every agent has a cryptographically attested identity with least-privilege permissions, propagated through OAuth to external systems; all actions go to an audit trail attributed to the agent, including in VMs the agent spins up. Execution runs in an Agent Sandbox, with all traffic through Agent Gateway, an AI network firewall that applies policies once across all agents. Cost controls include multi-model orchestration, Smart Routing that triages workloads to the cheapest adequate model, and real-time per-project spend caps in Cloud Billing Console that pause an agent when exceeded. Google notes per-token prices have dropped 98% since 2024 while enterprise volume has exploded.

On infrastructure, the AI Hypercomputer co-designs chips, network, data center, and software; the TPU 8i delivers 80% better price-performance than the prior generation. The model family includes Argon for reasoning, Flash for speed and volume, Omni for generative media, and Gemma for lightweight open-weights edge workloads. NASA’s Jet Propulsion Laboratory is already running Gemma on a satellite in orbit through its NAVI software, analyzing Earth imagery on board and sending summaries to ground operators.

Customer evidence spans scale and measurable outcomes. Google states that nearly 80% of all Google Cloud customers use its AI products, nearly 90% of the Fortune 100 use Gemini Enterprise, and nearly 500 customers each processed more than one trillion tokens in the last year. Reported results include Bradesco reducing document review time from one hour to five minutes with a 60% reduction in risk inconsistencies and over 10% financial efficiencies; Bunnings saving half a million hours of administrative work; Kmart and Officeworks raising conversion up to three times; Bloomberg Media lifting SQL query accuracy 63% by grounding in the Knowledge Catalog; Etsy cutting critical joins up to 60% after migrating three petabytes; Snap reducing diagnostic troubleshooting from 30 minutes to 30 seconds; Commerzbank cutting document review from 20 hours to one; Nokia cutting issue resolution up to 80%; Airwallex resolving customer issues 10x faster and reducing dispute case time 40%; DBS running deterministic chains of 70–80 specialized agents in production; NTT DOCOMO giving 450,000 hours back per year; Tata Steel cutting complaint turnaround 50%; Hitachi boosting productivity up to 400%; and Home Depot speeding phone resolution four-fold. The U.S. CDAO has made Gemini Enterprise available to 3 million service members, with more than 100,000 custom agents built.

Google closes with three commitments: a platform that runs anywhere and preserves model choice, a fully integrated stack from silicon to security, and customer ownership of proprietary inputs and outputs without requiring data or people to move.