Contents6
Until now, AI at work looked like a conversation: you ask a question, it answers, you decide. That model is giving way to another, where you hand over a goal and come back later to see the result. On 8 October 2026, Google Cloud gave the clearest example of it yet with Gemini agent. This article from our Business AI section looks at what this shift asks of a team of 20 to 500 people, beyond the announcement effect.
In short
Work agents no longer respond to instructions, they pursue goals, sometimes for hours.
Gemini agent creates its own sub-agents, each with its own identity: a sign that identity is becoming the central issue.
Four questions decide whether an agent is useful or dangerous: who it is, what it can touch, what trace it leaves, and who validates.
These questions are not specific to one vendor: they apply to every agent on the market.
What Google Cloud announced
According to Constellation Research, Gemini agent is a single agent for work, accessible from Google Workspace, Microsoft 365, Slack and other applications. Thomas Kurian, who runs Google Cloud, describes it as an agent that is given goals rather than instructions. It can plan, schedule tasks, react to events, and orchestrate processes that last for hours.
- 17 July 2025
OpenAI presents ChatGPT agent, which chains multi-step tasks with its own browser.
- 8 October 2026
Google Cloud launches Gemini agent, in early access for businesses.
- Late October or early November 2026
General availability expected, according to Constellation Research.
The most telling detail is not the duration. It is the way the agent organises itself: it can create a team of sub-agents, each with its own identity, role and email address. In the same announcement, Google also highlights agent identities, permissions, audit logs, isolation and spending caps. OpenAI had gone in the same direction with ChatGPT agent, and Microsoft offers its own agents in Copilot: the trend goes beyond a single vendor.
When AI starts delegating to other AI, the question is no longer what it can do, but on whose behalf it acts.
What changes for a team
A conversational assistant does nothing without you. An agent that runs for hours does things while you are elsewhere. That takes nothing away from the team’s role: it shifts its work from execution towards framing and control. Someone has to define the goal, set the limits, and review what comes out.
The benefits are real. A file that took a day of gathering information across five tools can be prepared overnight; monitoring, a follow-up, a data reconciliation can run without anyone watching the screen. But each of these gains assumes the agent has access to your tools, and that is where the difference between saving time and taking a risk is decided.
The four questions to settle before connecting
- Who is the agent?
It needs its own identity, distinct from an employee’s. An agent acting with someone’s account makes it impossible to know, later, who did what.
- What can it touch?
Permissions are cut to fit the mission: read-only by default, write access only where necessary, never administrator access “to go faster”.
- What trace does it leave?
Every action must be recorded in a readable log: which tool, which data, at what time, for which goal.
- Who validates?
Actions that commit the company (sending to a customer, paying, deleting, publishing) go through a person before they are carried out.
These four questions also apply to sub-agents. An agent that creates other agents itself must not be able to give them more permissions than it has. Our guide to AI governance in the enterprise explains how to divide these permissions and who is responsible for them.
Watch out
The shared password, a classic trap
The most tempting shortcut is to give the agent an employee’s login details. It is also the riskiest: the agent inherits all of their permissions, and its actions blend in with the person’s. An agent’s keys are kept in a vault, and the agent uses them without ever seeing them.
Cost, the other limit to set
An agent that runs for hours consumes for hours. Google understood this, announcing real-time spending caps and automatic model selection by task. For a team, the rule is simple: one budget per mission, a cap that stops the agent when it is reached, and a dashboard showing what each task cost. That is the principle of our AI at the right cost offer.
Our view
Autonomous agents will become a common tool, whichever vendor is chosen. The companies that get the most out of them will not be those that connect them fastest, but those that set the framework first: identities, permissions, logs, validation, budget. This framework is also what makes it possible to hand them more over time, as trust builds.
To understand how an agent reasons and acts in a loop, our article on agentic AI lays the foundations; to store its keys, see our secrets vault. Our guide to integrating AI in business puts these choices back in the order they are made. And if you want to know which mission to start with, an AI audit sprint maps your tasks and the permissions they require.
Frequently asked questions
What is Google Cloud’s Gemini agent?
It is an agent for work, announced on 8 October 2026, that is given a goal rather than an instruction, uses the company’s tools and can create sub-agents, each with its own identity. It can carry out processes that last for hours. It is first available in early access.
Can an AI agent replace a person in a team?
It takes on tasks, not a whole role. Someone in the team still has to define the goal, set the limits, review the result and validate the actions that commit the company. The team’s work shifts towards framing, control and decisions.
Should an AI agent be given an employee’s access?
No. An agent should have its own identity and permissions limited to its mission, stored in a secrets vault. Sharing a person’s login details gives it too many permissions and makes its actions impossible to tell apart from the employee’s.
How do you control what an agent does for hours?
Through a log of every action, a spending cap that stops it, and human validation points before sensitive actions. Reviewing the log at the end of a mission also lets you adjust its permissions for the next one, removing those it did not use.
Sources
- Constellation Research, Google Cloud launches Gemini agent (new tab)constellationr.com
- OpenAI, Introducing ChatGPT agent (new tab)openai.com
Constellation Research, Google Cloud launches Gemini agent
Read on constellationr.com(new tab)Jim · founder of Brain Plus AI
Also in business ai
Practical guide12 min
Enterprise AI governance: who can access what
An AI assistant reads whatever your access rights let it read, sharing mistakes included. Here is how to split your knowledge bases by department, keep a common core and decide who writes to the shared memory.
Understand10 min
Agentic AI: how an agent's loop works
An AI agent does not do everything in one go: it runs in a loop, gathers context, acts, checks and starts again. Understanding that loop means knowing where to put the controls, the limits and the human review.
Understand12 min
Secrets vault: your AI uses your keys without seeing them
Passwords sent by email, API keys forgotten in code, AI agents that could copy everything: these are the leaks we come across most often. And how a vault in your name lets your AI use your keys without ever reading them.