Google Launches Gemini Agent: One AI Agent That Runs Your Work Across Models
Google Cloud launched the Gemini agent on Oct 8: a universal work agent that takes objectives, spawns sub-agents, routes jobs across models, and finishes tasks inside Workspace.
At its Gemini at Work 2026 event on October 8, Google Cloud introduced the Gemini agent: a single, universal AI agent for enterprise work. Instead of answering prompts one at a time, the agent takes an objective, plans the steps, picks the right tools and models, connects to your business systems, and returns finished work inside Gmail, Docs, Sheets, Slides, or your developer environment.
What the Gemini agent actually does
Google Cloud CEO Thomas Kurian framed it as objectives, not instructions. Tell the agent what outcome you want and it plans the work itself: researching a market, building a financial model in Sheets, and drafting the presentation from the findings, all from a single prompt. It can answer questions, handle knowledge work, create media, and write and run code.
The agent runs in the cloud, so jobs keep going for hours or days after you close the laptop. It spawns temporary sub-agents that work in parallel or in sequence, coordinating workflows while you do other things. Progress shows up in a tasks inbox with the reasoning, delegated sub-tasks, loaded skills, and code written visible to the user.
Memory, model routing, and governance
The system keeps four kinds of memory: session memory for the task in front of it even across days, semantic memory built from documents and people, procedural memory storing how jobs get done, and episodic memory recording everything it has done before. Skills are modular prompts stored in a shared company registry, and connectors cover Workspace, Microsoft 365, Slack, Jira, Confluence, Git, BigQuery, Databricks, Postgres, Snowflake, and any Model Context Protocol server.
Notably, the agent runs each job on the best-fit model: Gemini and Anthropic's Claude families at launch, with other private and open models planned. Built-in cost controls, security measures, administration features, and governance capabilities are designed for enterprise customers. The agent is in private preview for Gemini Enterprise customers, with general availability expected by the end of October.
What this means for your business
Three practical takeaways. First, evaluate agents on finished work, not answers. The vendors are converging on the same pitch: give an objective, get completed work. When you pilot these systems, define success as a finished deliverable (a model built, a deck drafted, a ticket resolved), not response quality.
Second, model-agnostic routing is becoming standard. Google routing jobs to Claude where it fits best signals that the era of one-company model lock-in is fading. Build your AI procurement around interoperability and cost-per-job, not loyalty to a single vendor's model family.
Third, governance travels with capability. Cost controls, audit trails, and admin oversight are shipping alongside the agents, not after. If your team experiments with autonomous agents before these controls exist in your environment, keep them read-only on company systems until the guardrails catch up.