Gemini Agent Arrives | The AI That Takes Your Entire Job Off Your Hands
機械翻訳 / Machine-translated
A
AI Friends
16時間前・17分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever thought, "I wish I could hand off an entire task to AI — like conducting market research, putting it into a table, and building a presentation"? Google's "Gemini Agent," announced on October 8, 2026, is an AI aimed at making exactly that possible. Here's a plain-language explanation of what it can do and how it differs from the competition.
At its October 8 event "Gemini at Work 2026," Google announced Gemini Agent. The official Google Cloud blog positions it as a "universal agent for work."
An agent is an AI that plans and carries out assigned tasks on its own. Unlike a chat AI that simply answers questions, an agent receives a goal, figures out the steps, and sees the work through to completion.
According to the official blog, users only need to hand over a goal, not a set of instructions. Answering questions, generating images, and writing and running code can all be requested through a single input field and a single API (the interface that connects software applications).
It is a feature being added to the enterprise service "Gemini Enterprise," and according to GIGAZINE's coverage, it is currently in limited preview for select businesses. No general availability date has been announced.
Imagine a sales planning manager who has been asked to conduct market research for a new product. Until now, that meant searching for data, analyzing it in a spreadsheet, and then compiling everything into a presentation.
With Gemini Agent, research, analysis sheets, and presentation slides can all be handled in a single, continuous flow. It operates within apps like Gmail, Drive, Docs, Sheets, and Calendar.
Even scheduling meetings requires no name or email address entry — the agent infers attendees from chat participants and past exchanges.
Tasks are executed in the cloud (computers on the internet). Work that takes hours or even days continues even after you close your notebook computer.
It is accessible from the web, iOS, Android, Windows, and Mac, as well as from Microsoft 365 and Slack. The same memory is shared across all devices.
There are four types of memory: memory of the current task, memory of knowledge gained on the job, memory of how tasks are done, and a history of past work.
The goal is that you never need to re-explain context. Simply saying "continuing from the project I asked about before" is enough.
Additionally, temporary sub-agents can be created for specific jobs and run in parallel.
Another eye-catching feature is the "colleague agent." Simply describe a role in words, and a new team member is created — complete with a dedicated email address, calendar, and Drive.
You can assign work by mentioning the agent in Google Chat or tagging it in a document comment. A record of the work also remains in the version history.
Perhaps surprisingly, models from outside Google are also available. According to the official blog, users can choose between Gemini models and Anthropic's Claude models depending on the task.
The system automatically routes complex tasks to high-performance models and simpler tasks to more affordable ones, balancing accuracy and cost. Google's model lineup includes families such as Argon (for advanced reasoning) and Flash (for speed).
Plans to expand to other private and open models are also in the works.
When handing work off to AI, perhaps the biggest concern is not knowing what it might do on its own. The official blog addresses security through four questions:
For example, a rule such as "do not open confidential documents" can be written once and applied to every agent across the organization.
On the financial side, a key feature is the ability to set a spending cap for each project. When the cap is reached, the agent stops working and can be resumed with a single button press — preventing any unpleasant billing surprises.
According to The Decoder, Google announced Gemini Enterprise on October 9, 2025, describing it as a response to Microsoft Copilot and ChatGPT Enterprise. The new agent is a continuation of that direction.
That said, it's difficult to make definitive claims about which is better. Many comparison articles come from sources close to the vendors, and the metrics used aren't standardized. Broad tendencies are as follows:
Pricing is reported as starting at $21 per user per month for Gemini Business, and $30 per user per month for Gemini Enterprise Standard/Plus. These are prices for Gemini Enterprise as a whole, however. Pricing for Gemini Agent on its own is not listed on the official blog.
The official blog includes several Japanese case studies.
SOMPO is highlighted as using over 10,000 custom agents across 34,000 employees, cutting model development time from one week to one day.
NTT Docomo reportedly reduced the time needed to find information from two weeks to near-instant, reclaiming 450,000 hours annually. Hitachi has seen productivity improvements of up to 30%, and Marubeni is listed as an Asia-Pacific adopter. All figures, however, are self-reported by each company.
There are also open questions. Japanese language support and the timeline for availability in Japan were not clarified in this announcement. It is not yet known when Japanese companies will be able to start using it.
What about individuals and small businesses? Since the current announcement is aimed at enterprises, it is not something anyone can use right away. However, the approach — where you simply state a goal in a single input field and the work gets done — is expected to spread to other AI services going forward.
A. It is in limited preview for select businesses, and no general availability date has been announced. It is not something individuals can start using today.
A. Pricing for the agent itself has not been disclosed. Gemini Enterprise is reported to start at $21 or $30 per user per month.
A. Its distinguishing features are its integration with work applications and its enterprise-grade management capabilities. The ability to choose Claude in addition to Gemini is also a difference.
A. Each agent is given a unique ID and specific permissions, and all operations are logged. Agents run in isolated execution environments (sandboxes), and communications are monitored by Agent Gateway. That said, establishing operational rules remains the responsibility of the implementing company.
As a next step, try writing down one task your organization repeats every month — research or document creation — and consider it a candidate to hand off to AI.
This article is a cross-post from AI Friends.
