Google brings agentic AI to Gemini, starting with businesses

Google is turning Gemini into an agent that can plan, carry out tasks and work across business apps and systems, according to TechCrunch. It can hand pieces of work to subagents, draw on multiple AI models, and gets its own workplace identity, complete with an email address. Businesses get it first.

That is a shift from a chatbot you ask questions to a worker you assign things to. For companies already living in Google's productivity stack, it could take real busywork off people's plates. It also means a piece of software now holds an account inside the company, with whatever access that account is given.

We think starting with businesses is the right call. Enterprises have admin consoles, audit logs and people whose job is to say no. That is a better first home for an agent that acts on its own than a consumer inbox with no one watching.

The workplace identity is the detail I keep coming back to. Giving the agent its own email address is honest in one sense: actions get attributed to the agent, not smeared across a human's account. That helps audits. But it also creates a new kind of employee that never sleeps, can be messaged by anyone, and can be manipulated by whatever lands in its inbox. Prompt injection stops being a lab curiosity when the target has an address and permissions.

Subagents and multiple models add capability and add opacity. When a task is split, routed and reassembled, the question of who decided what gets harder to answer. Businesses will need to see that chain, not just the final result.

The upside is real. Plenty of office work is coordination: chasing approvals, pulling numbers from three systems, drafting the first version. An agent that does that reliably is useful, and Google has the distribution to make it ordinary fast. The risk is that ordinary arrives before the controls do. This has been a year of agents going past the boundaries they were given, and an agent with a corporate identity is exactly the kind of system where that matters.

Our read: promising, and the identity model is a smart foundation, as long as permissions default narrow and every action is logged. I'm watching what admins can actually restrict on day one.

Google, Microsoft, OpenAI and others are all pushing agents that act inside workplace software rather than just answering questions. The competition is increasingly about who controls the systems of record those agents can touch.