Google Gives Gemini an Agent Mode for Business, Complete With Its Own Work Email

Google is rebuilding Gemini around agentic workflows, giving the assistant the ability to plan multi-step jobs, carry them out across business applications, and operate inside corporate systems rather than waiting on prompts. The company is starting with business customers, where repetitive, cross-app work is the clearest target. Each agent also receives its own workplace identity, including an email address, so it can be assigned tasks and contacted like a colleague.

The architecture Google describes is hierarchical. A top-level agent breaks a goal into steps, then hands pieces off to subagents that run in parallel or in sequence. The system can also route work across more than one AI model, meaning a single task might draw on different Gemini variants depending on what each step requires. That is a departure from the single-model, single-conversation pattern that has defined most assistant products since ChatGPT launched.

Handing an agent an email address and login credentials changes the security calculus. An autonomous process that can read mail, touch internal systems and act without a human in the loop needs permission boundaries, audit trails and a way to revoke access quickly. Google's pitch to IT buyers will hinge on whether those controls are mature enough to satisfy security teams that have spent the past two years restricting how generative tools touch corporate data.

The move puts Google in direct competition with Microsoft, which has been pushing Copilot agents through its own productivity stack, and with a wave of startups selling agent frameworks to enterprises. Google's advantage is distribution: Gemini already sits inside Workspace, Android and Google Cloud, so an agent can reach users without a separate install. Its challenge is the same one every vendor faces, which is proving that agents finish real work reliably instead of stalling halfway through a task.

For developers and platform teams, the interesting question is how open the agent layer will be. If Google lets third parties register tools, models and subagents, Gemini becomes a runtime rather than a product. If it stays closed, enterprises will treat it as another feature to evaluate alongside Microsoft's and Anthropic's offerings. Google has not detailed pricing, availability dates or which model families the agent can call.

The broader shift here is that assistants are becoming infrastructure. Once an AI system has an identity, permissions and the ability to delegate, it stops being software you talk to and starts being a worker you manage. The next few quarters will show whether enterprises are ready to give it a seat.