At its Gemini at Work event on 8 October, Google Cloud introduced a Gemini agent that takes a goal rather than a prompt. Thomas Kurian described it as: “You give it objectives, not just instructions.” It plans the work, coordinates sub-agents, connects to Gmail, Docs, Drive, Sheets and Calendar, and can keep working for hours or days on a single objective while retaining memory. Writing and running code is one of its core jobs.
It is in private preview for select customers through Gemini Enterprise and Google Workspace. There is no general availability date and no published pricing. Google also published no coding benchmarks, so any claim that it beats Claude Code, Codex or Cursor is unverified.
What is genuinely new
The pattern, not the product. Every major vendor is converging on the same shape: a long-running agent, built-in memory, sub-agents and cost controls, sold as a workplace tool rather than a developer tool. Google says the agent “chooses the best model for the job” and has built-in cost controls, which tells you where buyer anxiety about AI bills now sits.
What buyers should not conclude
- That you no longer need custom software. A general agent inside Workspace is good at generic work. It does not know your pricing rules, your compliance obligations or your customers.
- That a preview is a plan. Private preview features change, get repriced or vanish. Do not hang a roadmap on one.
- That autonomy equals reliability. An agent that works for days can also be wrong for days. Approval points and logging matter more as run times grow.
So what
Treat this as a signal to separate two decisions: which work a general-purpose agent can absorb cheaply, and which workflows are specific enough to your business to justify a purpose-built product with approvals, audit trails and the freedom to switch models. If you are scoping the second kind, BuildApps designs AI products with those controls built in, as part of AI-assisted development. To talk through a project, start a project.