Google announced a universal work-focused “Gemini agent” at its Gemini at Work 2026 event. After the user gives it business context, the agent can answer queries, create images, write and run code, among other activities. Moreover, it can plan, use tools and connect with your workflows to render results in a document, inbox or more.
Gemini agent can choose between models (to deliver the best output at minimum cost) while providing enterprise-grade security. It can be accessed from Android, iOS, web, Windows, Macs and channels such as command line, Microsoft 365, Slack, etc. Plus, the Gemini agent can be integrated into third-party apps.
“Gemini can function as your personal assistant or as a team member, where it works on behalf of a group of people,” a Google Cloud blog post read. It can also create sub-agents or coworker agents and assign them responsibilities and a custom email ID. These agents can connect to a company’s tools and software to accomplish tasks while continuing to learn via “Skills” training using the given instructions.
The agent keeps four kinds of memory: session memory for a specific task, semantic memory as a base of knowledge learned while executing multi-step tasks, procedural memory to remember the task’s process and a combined episodic memory of previous tasks.
Google also highlighted capabilities of Gemini in Workspace as a personal assistant that can delegate business tasks to AI when a user wants. Further, the assistant will also be optimised for domain-specific use cases (such as of data and ML engineers). Plus, specialised versions of Gemini are now in preview for legal and financial service sectors and are expected soon for government, retail and healthcare domains.
For optimised data storage for businesses, Google announced Smart Storage and Knowledge Catalog to help the AI learn organisation-specific knowledge and terms. Businesses can implement limits for projects to keep spending in check. Meanwhile, Google also highlighted that its latest TPU 8i can deliver an 80 per cent better price-to-performance ratio than its predecessors.