Gemini agent: Google's single agent for work, code and Workspace

Google wants one agent instead of a dozen. At its Gemini at Work 2026 event on 8 October, Google Cloud announced the Gemini agent, a "single, universal agent for work" that answers questions, does knowledge work, creates images and media, and writes and runs code, from one prompt box and through one API. The announcement is long; here is what matters if you build or code.

What the Gemini agent is

Google describes the following capabilities in its announcement; this is not an independent product test.

• You give it goals, not steps. It plans the work, picks skills and tools, connects to company systems and brings back finished work in documents, the inbox or the developer environment.
• It runs in the cloud and keeps going. Memory and context stay the same on every device and channel: web, iOS and Android, Windows and Mac desktops, the command line, Google Workspace, Microsoft 365 and Slack. It can also run headless inside other apps. Work that takes hours or days continues after you close the laptop.
• It brings helpers. The agent spins up temporary sub-agents for multi-step jobs, and "coworker agents" with their own identity, their own @agents.company.com email address and their own storage, which colleagues add to a chat or mention in a document.
• The model is a setting. Google says the agent runs each job on the model that fits, across the Gemini family and, today, Claude models from Anthropic, with other models to follow.
• Tools, skills, memory. It connects to tools such as Git, Jira, Slack, BigQuery and Postgres and to any MCP server inside or outside the company network. Teams publish tools and skills to shared registries, and the agent keeps four kinds of memory: for the session, for facts, for procedures and for what it did before.

Guardrails and costs

Google says each agent gets its own identity with least-privilege permissions and an audit trail. Its announced controls include an Agent Sandbox with a network boundary and Agent Gateway, described as an AI network firewall enforcing company policy. Google also describes routing to the cheapest suitable model and real-time spend caps per project that pause the agent when reached. Versions for financial services and legal work are in preview. Google's own figure for its new TPU 8i system is 80% better price-performance than the previous generation.

What the announcement does not say: when the Gemini agent will be generally available, and what it costs.

Our take

The bet behind it: the agent is the product, and the model underneath is interchangeable. It is notable that Google builds that bet with Claude models next to its own. For developers, the command line, MCP and the skills registry are the interesting parts. For companies, spend caps and an agent firewall decide whether anyone switches these things on. Without a date and a price, it stays a promise for now.

In AyCode

AyCode runs Google's Gemini models through Antigravity, Google's coding agent, next to Claude Code, Codex and others (all agents). The Gemini agent itself is a Google Cloud product and does not run in AyCode. The idea of coworker agents has a counterpart in AyCode, though: Aylinge are agents with a name, a brief, a memory and routines, each on the provider you choose (docs).