Google Cloud CEO Thomas Kurian unveiled the Gemini agent at Gemini at Work 2026, positioned as a single universal work agent: users provide goals rather than instructions, and it completes knowledge work, content and image media generation, as well as coding and execution.

The agent can be accessed via web, iOS, Android, Windows, and Mac, and can also operate through command line, Workspace, Microsoft 365, Slack, and connect to third-party applications in embedded or headless modes. Execution continues in the cloud, allowing tasks to persist across hours and days even after closing the laptop; memory, context, and personalization maps are unified across devices without needing to re-explain. It can temporarily generate sub-agents with independent identities to coordinate in parallel or serially; it can also become a long-term colleague agent with a dedicated email, storage, and directory identity, only reading shared context.

The underlying model is separate from the agent. Tasks can switch between Google’s Gemini series and Anthropic’s Claude, with simple tasks not requiring the largest models, while advanced tasks utilize larger models to match quality and reduce costs. Cost control includes multi-model orchestration, intelligent routing, and real-time spending limits. Governance provides identity and policy management, authorization and permissions, security sandboxes, and network gateways. Connectors cover Confluence, Teams, Slack, Git, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Postgres, Snowflake, and model context protocol services both inside and outside enterprises.

Kurian stated the scale: nearly 500 Google Cloud customers processed over 1 trillion tokens each in the past year; nearly 80% of Google Cloud customers are using its AI products; and nearly 90% of the Fortune 100 are using Gemini Enterprise. Google CEO Sundar Pichai mentioned at the event that Gemini has over 1 billion monthly active users. The agent is directly embedded in Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar, maintaining the same set of memory, skills, and controls. New natural language data analysis skills have been added, providing industry tools and knowledge to financial services and legal teams.

In early tests, the sports brand On accelerated its market launch using dynamic model selection. Shopify mixed cutting-edge models across millions of merchants. PayPal routes 10 million multi-model requests weekly, stating that the rollout of new models has reduced from weeks to minutes. Bradesco reduced document review time from 1 hour to 5 minutes, with risk inconsistencies dropping by 60% and financial efficiency improving by over 10%. Orange Spain deployed over 1,000 custom agents. Japanese insurance company SOMPO built over 10,000 agents among its 34,000 employees, reducing new model development time from 1 week to 1 day. Bunnings' internal agent saved employees 500,000 hours of administrative time.

The buyer is the enterprise seat and token budget already running on Google Cloud, while the seller is the models and tools routed by task. The event is driven by the agent transforming from a chat box into a delegable work entry point. Funds flow through intelligent routing to the cheaper or more suitable side of Gemini and Claude, with Workspace, Microsoft 365, and Slack becoming distribution channels rather than replaced interfaces. Beneficiaries include cloud seats, tokens, and selected models; those under pressure include single model seats and point agents that require separate system integration and governance.

Source: Public Information