What Is Google's Gemini Agent? Gemini at Work 2026 Explained
The short answer
The Gemini agent is Google Cloud’s attempt to replace a dozen separate AI assistants with one agent that takes objectives instead of instructions. Google announced it on October 8, 2026 at Gemini at Work 2026. As of October 12, 2026 it is in private preview, with wider availability promised “soon” for Workspace customers on select Business and Enterprise plans, and no published price.
What it does, in Google’s words: it “answers your questions, handles your knowledge work, creates your images and media, and writes and runs code — all in a single agent and a single API.”
What is new about it
| Capability | What Google announced |
|---|---|
| One agent, one API | Chat, autonomous tasks and code generation from one interface; can run headless inside third-party apps |
| Runs anywhere | Web, iOS, Android, Windows, Mac; Workspace, Microsoft 365, Slack and command line; @Gemini in Gmail, Docs, Sheets, Slides and Chat |
| Long-running work | Runs in the cloud, so tasks that take hours or days continue after you close the laptop |
| Sub-agents | Creates temporary, job-specific sub-agents, each with its own identity, for parallel and sequential steps |
| Coworker agents | Persistent team-role agents with their own Workspace account, email address, calendar and Drive |
| Model choice | Routes each job to the best-fitting model — Gemini models and Anthropic’s Claude models today, other private and open models later |
| Memory | Session, semantic, procedural (including skills it writes itself) and episodic memory, per 9to5Google |
| Industry versions | Preview for Financial Services and Legal; Government, Healthcare and Retail “coming soon” |
Governance and cost controls
Google framed security around four questions — who the agent is, what it may access, what it did, and what traffic it sends — and answered them with:
- Identity: every agent gets a cryptographically attested identity, governed “like an employee”, stamped into its logs and into any VM it starts.
- Permissions: fine-grained, role-based access approved by security admins; identities propagate to external systems through standards such as OAuth.
- Audit: every action is attributed to the agent, not to a person.
- Agent Sandbox and Agent Gateway: tasks run in a sandbox with its own network boundary; all traffic passes through Agent Gateway.
- Smart routing and real-time spend caps: cheaper models for simple loops, and a hard project budget in the Cloud Billing Console that pauses the agent when reached.
How it compares
The Gemini agent enters a category that already has OpenAI’s ChatGPT Work and Anthropic’s Claude Cowork; our ChatGPT Work vs Claude Cowork vs Gemini agents comparison covers that decision. Two things set Google’s version apart on paper: it can run Claude models inside a Google product, and coworker agents get full Workspace identities rather than borrowing a human’s account. For the governance layer across vendors, see neutral AI agent oversight platforms.
What is still unknown (October 12, 2026)
- Price — no list price for the agent or for coworker agents.
- General availability date — Google says “soon”; Futurum reports end of October 2026, starting in North America.
- Which plans include it — “select Business and Enterprise plans” has not been mapped to named SKUs.
- Benchmarks — Google published customer results, not task-completion benchmarks.
Model prices for the Gemini and Claude families are on our current API prices page.
Last verified: October 12, 2026.