Google launches Gemini Agent, an office AI that can adopt a work account and call Anthropic's Claude

Google Cloud has unveiled Gemini Agent, a general-purpose office agent built to run long tasks across applications, in a lengthy post titled "Welcome to Gemini at Work 2026: Introducing the Gemini agent" signed by Google Cloud chief executive Thomas Kurian.

The company frames the product around a simple principle: you give it objectives, not instructions. Instead of laying out steps, a user states the outcome they want and the agent plans the work itself. In one example, it researches and organizes material, analyzes data, builds a financial model in Google Sheets and produces a slide deck in Google Slides.

Google says the core capabilities include a unified entry point, persistent cloud execution, cross-platform tool calls, multi-agent collaboration, event triggers and scheduled tasks. When a job is large enough, the agent can convene several sub-agents to work in parallel.

Two features stand out. The first is identity: under a concept Google calls Coworker Agent, a company can create a long-lived Gemini identity with its own Google Workspace account, email, calendar and Drive storage, listed in the staff directory and reachable in Google Chat or via an @-mention in a document. It acts under its own name and leaves a record, which Google pitches as a management and audit benefit.

The second is model choice. Gemini Agent can pick its underlying model, switching dynamically between Google's own Gemini family and Anthropic's Claude, with plans to add more closed and open models. Google calls this multi-model orchestration and pairs it with Smart Routing, which weighs effectiveness, speed and cost so that simple jobs run on cheaper models and complex ones on stronger models. Context, skills and enterprise data are meant to survive a change of underlying model.

Google also describes four kinds of memory: session memory for a task's live state, semantic memory for company knowledge, procedural memory for how a job should be done, and episodic memory for what the agent has done before. The agent can even write its own Skills to save a method for reuse. Google likens the experience to a new employee gradually getting to know you, your tools and your team.

The launch lands in a crowded field. OpenAI recently introduced Dots, an agent that can run around the clock on its own cloud computer, while Meta's Muse leans toward consumer life tasks such as shopping and booking travel. Google's pitch is narrower: pulling agents deep into existing enterprise workflows through Gmail, Docs, Sheets and Calendar, with model choice, enterprise data connections, security auditing and cost controls bundled in.

Alongside Gemini Agent, Google said Gemini is moving deeper into Google Workspace, opening up Skills, tools and enterprise connectors, and adding agents specialized for data analysis, finance and legal work, together with new agent security governance and cost controls.