Who manages an agent that has its own mailbox? Google gives Gemini coworker agents a Workspace account

On 8 October 2026, at Gemini at Work 2026, Google Cloud presented the Gemini agent, a single agent for work. One of its forms concerns IT directly: the coworker agent. It gets a Workspace account, an email address, a calendar, a Drive and a place in the company directory. For IT, this is no longer a tool. It is one more account to manage.

What Google announced

These points come from Google Cloud's press release, which we read in the copy distributed by Zawya.

• A role, not a task. "You simply describe the role you need, and Gemini creates it." The coworker agent has a defined, persistent role, like a team member.
• A Workspace account. Email address (of the form @agents.company.com in Google's example), calendar, Drive and presence in the company directory.
• Its own identity. The agent acts under its own identity, not yours. Google writes that this identity is "cryptographically attested and governed like an employee".
• A log in its name. Every action is written to an audit trail and attributed to the agent, not to a person.
• A sharing scope. The agent "sees only what you share with it". Permissions are approved by the organisation's security administrators.
• Several models. The agent orchestrates Google's Gemini models and, already, Claude models.

Established, reported, not published

Point	Status	Where it comes from
Own account, email, calendar, Drive, directory entry	Established	Google Cloud press release
Actions logged in the agent's name	Established	Google Cloud press release
Spend cap per project that pauses the agent until someone resumes it	Reported	Decrypt, Reworked, Implicator. Google's text only mentions "built-in cost controls".
No extra charge where Gemini Enterprise is available	Reported	A Google spokesperson to VentureBeat, according to Reworked and Implicator
Private preview, general availability expected late October or early November	Reported	Implicator, citing executives speaking to analysts
Price and general availability date	Not published	Google has published neither
Whether every agent gets an account, or only coworker agents	Not established	Implicator notes that Google has not said

What it changes for you

An agent with a mailbox and a directory entry is an account in its own right. It needs what you require of any account: an owner, a scope and a leaving procedure. The log in the agent's name says what the agent did. It does not say which person answers for it. That answer is yours to write down.

The agent's real scope is the state of your sharing. If it "sees only what you share with it", any folder open to the whole company is readable by it. Existing oversharing on your Drive becomes its playing field.

An agent that writes to third parties from its own address raises the question of informing people. Article 50 of the AI Act has applied since 2 August 2026 (see what the AI Act requires of companies that deploy AI). It requires providers to design systems that interact directly with people so that those people know they are dealing with an AI, unless it is obvious. The European Commission cites AI agents among its examples. How this applies to an email sent by a coworker agent, and what falls to the company using it, is not settled here: check it with your legal team. Nothing stops you from saying it in the signature in the meantime.

The gap is not specific to Google. In a survey released in early October by SailPoint, a vendor of identity security tools, of 340 identity, IT, cybersecurity and risk leaders: 79% of organisations run AI agents in production, 2% use identity tools built to govern them. 57% say they are confident about meeting regulatory requirements, 43% say they are prepared to produce verifiable evidence in an AI-related audit. SailPoint sells these tools: read these figures as a vendor's.

The danger, the advice

The danger	The advice
A manager creates a coworker agent in a few minutes. Nobody in IT knows it exists, and nobody is named as its owner.	Decide before the first activation who can create an agent, and record each one in your AI register with a named owner on the day it is created.
The agent "sees only what you share with it", and your Drive already shares too much. Folders open to the whole company become its working material.	Review company-wide shares and public links before activation. The agent's scope is your sharing hygiene.
The log attributes every action to the agent. During an incident or an audit, nobody can say which person answers for it.	Keep, next to each agent, the name of its owner and the person who approved its permissions. Read its log at fixed intervals.
The agent writes to customers or suppliers from its own address, and they do not know they are dealing with an AI.	Say it in the signature, and have your legal team check what Article 50 of the AI Act requires in your case.
A project ends, the agent stays: active account, open mailbox, access still granted.	Write a leaving procedure for agents: suspension, handover of mailbox and files, revocation of access.

To decide before switching it on

1. Who can create a coworker agent. Restrict it to a group, not to every manager.
2. Who owns each agent. One named person per agent, recorded in your AI register on the day it is created, as for any AI agent in production.
3. What it can see. Review company-wide shares and public links before the first agent arrives.
4. How to stop it. A written procedure: suspend the account, hand over mailbox and files, revoke access. The same as for someone leaving.
5. What it can spend. A cap per project if your console offers it, otherwise a monthly review of consumption.

What a solid team does differently

It ties each agent to a person, reads its action log at fixed intervals, and treats the creation of an agent like an arrival and its deletion like a departure. It applies the same grid to agents on other platforms: the vendor changes, the questions stay the same.

This set-up needs people who know Workspace and Google Cloud from the inside. Our Google Cloud experts can frame these choices with your teams. For the governance itself (register, owners, evidence), see 56North.