When an AI agent gets its own inbox, what has it earned?

On October 8, at its Gemini at Work 2026 event, Google Cloud introduced the Gemini agent, which it calls a single agent for work. The detail that stands out is the coworker agent. A team describes the role it needs, and the agent receives its own Workspace account, with an email address, a calendar, Drive storage and a place in the company directory. This matters because it changes how an agent sits inside a company. It is no longer a feature inside someone's tool. It is set up more like a colleague.

What Google announced

According to Google, a coworker agent acts under its own identity rather than the identity of the person who set it up. It sees only what people share with it, and access follows the sharing rules a team already uses. Work can run for hours or days and keeps going after the person who started it closes their laptop. The agent can also split a job into parts and hand them to temporary sub-agents, each with its own identity.

On control, Google says each agent gets fine-grained, role-based permissions approved by the organisation's security administrators. Every action is written to an audit trail and attributed to the agent, not to a person. Administrators can write a policy once and have it apply to every agent in the company. Google says it is bringing the agent to businesses first.

Identity answers who, not how well

These are sensible foundations. Giving an agent its own identity solves a real problem. When an agent acts with a person's account, nobody can tell afterwards whether the person or the software sent the email or changed the file. A separate identity and an audit trail fix that.

But a record of who did something is not evidence that it was done well. An audit log will show that an agent updated a spreadsheet model on Tuesday. It will not show that the formula in row forty was wrong, or that the summary it sent to a client left out the one caveat that mattered. Permissions decide what an agent may touch. They say nothing about how good its work is once it touches it.

An identity tells you who did the work. Only review tells you whether the work was right.

Access should follow evidence

In most companies a new hire does not get the same access on day one as a ten-year veteran. Trust is extended in steps, as people show they can handle the work. The same idea fits agents with accounts of their own. The useful question is not whether an agent can do a task. It is whether it has shown, on this team's real work, that it does the task correctly.

In practice that suggests a few habits for any team setting up coworker agents:

• Write down what a correct result looks like for each task before the agent starts it.
• Start the agent on tasks where a mistake is cheap and easy to spot, and have someone who knows the work check the output.
• Widen its permissions one task at a time, only when the checks show it is consistently right.
• Treat every error a reviewer catches as a new test case to rerun after each model or prompt change.

None of this replaces the controls Google describes. It sits on top of them. The audit trail becomes far more useful when each entry can be matched to a judgment about whether the work was correct.

Long tasks raise the stakes

Work that runs for days, split across sub-agents, is where small errors travel furthest. A wrong assumption early in a long task can shape everything that follows, and by the time a person looks, the output may be polished and confident. General benchmarks do not catch this, because they rarely look like a particular team's work. The checks that catch it are built from that work: the real documents, the real edge cases and the mistakes the team has already seen.

What this means in practice

Agents with their own inboxes are a reasonable step, and the identity and logging that come with them make agents easier to govern. Teams should still treat a new coworker agent the way they would treat any new colleague. Give it a clear standard, check its work with people who know the job, and grant autonomy task by task as the evidence builds up. At SoReliable, that is the gap we focus on: turning expert judgment into a measured record of what an agent can be trusted to do.